“Is AI actually going to fix health inequity—or just give nicer tools to people who are already well served?”

In earlier Electric Sheep posts, I wrote about why I co-write with an AI assistant and how multimodal data—blood, breath, sleep, steps—could reshape chronic disease care.

In this post, I want to look at where AI is colliding with inequity in the real world—through FemTech (women’s health), AgeTech (aging and long-term support), and what I’ll call EquityTech: technology built first for people at the margins of health systems.

The core question throughout:

Who actually benefits from these systems—and who still gets left out?

1. Where inequity stands today

The U.S. Centers for Disease Control and Prevention (CDC) defines health equity as a state in which everyone has a fair and just opportunity to attain their highest level of health. That requires removing obstacles such as poverty, discrimination, and lack of access to good jobs, education, housing, and healthcare. (CDC – Health Equity in Chronic Disease)

The Agency for Healthcare Research and Quality (AHRQ) has been tracking this for more than 20 years. Its 2023 National Healthcare Quality and Disparities Report concludes that disparities by race, ethnicity, income, and geography remain persistent and substantial across prevention, diagnosis, and treatment for chronic disease—there is no broad pattern of gaps closing on their own.

Chronic diseases such as diabetes, hypertension, cardiovascular disease, sleep disorders, and depression fall heaviest on:

• Black, Hispanic/Latino, and Native communities

• Low-income populations

• Rural areas and medically underserved urban neighborhoods

These are also the groups least likely to have consistent access to early diagnosis, specialty care, and longitudinal support. Equity here is not abstract; it shows up in who loses kidneys, who has strokes in their 50s, and who dies decades earlier than they should.

We can frame these gaps along a care cascade:

1. Screening and triage – Who is identified as “at risk”?

2. Diagnosis – Who receives timely and accurate diagnoses?

3. Treatment initiation – Who starts appropriate therapy?

4. Adherence and follow-up – Who can sustain treatment with support?

5. Timely escalation – Whose deterioration triggers action before a crisis?

Failures occur at every stage, driven by system design, not just individual choices.

A few recurring drivers:

• Resource scarcity. When clinicians and care managers are scarce, algorithms are used to prioritize who gets outreach. If those algorithms are trained on biased data, inequity is amplified.

• Biased measurement. Even basic tools are less accurate for some groups. Pulse oximeters, for example, systematically overestimate blood oxygen saturation in many patients with darker skin, leading to missed or delayed detection of hypoxemia. (NEJM – Racial Bias in Pulse Oximetry)

• The digital divide. Telehealth and remote monitoring presume reliable internet access, devices, language support, and digital literacy. Without explicit planning, digital health can improve access for already-advantaged patients and leave others further behind.

On top of this, we are now introducing Large Language Models (LLMs) into clinical communication, patient education, and decision support. A 2025 systematic review of LLMs in clinical medicine highlights strong performance on some narrow tasks, but also substantial variation and clear risks related to bias, explainability, and inappropriate recommendations. (PubMed – LLMs in Clinical Medicine)

To state the situation simply: if we do not explicitly design and monitor for equity, AI systems will add inequity to already unequal care.

That’s the baseline on which AI now operates.

2. Where AI can help—if we design it for equity

Despite the risks, there is credible evidence that AI-enabled tools can improve equity when they are intentionally designed and evaluated with that goal.

2.1 Expanding access to evidence-based care

Some digital health products function primarily as access multipliers—they deliver proven interventions to people who would not otherwise receive them.

Digital CBT-I for insomnia.
Sleepio, a fully digital cognitive behavioral therapy for insomnia (CBT-I) program from Big Health, uses AI to personalize content and pacing. The UK’s National Institute for Health and Care Excellence (NICE) recommends Sleepio as a cost-saving primary-care option for adults with insomnia or insomnia symptoms, especially where access to therapist-delivered CBT-I is limited.
(NICE briefing)

Remote postpartum blood-pressure monitoring.
For women with hypertensive disorders of pregnancy—a major equity issue, given higher maternal mortality in Black and low-income populations—remote monitoring programs combine a blood-pressure cuff, mobile app, and protocol-driven alerts. A 2023 multicenter study found that remote postpartum blood-pressure monitoring significantly increased timely BP assessment and detection of severe hypertension compared with office-based surveillance alone.
(Study – Remote Monitoring for Postpartum HTN)

AI-enabled lifestyle and chronic-disease coaching.
Lark Health uses conversational AI, connected devices, and curricula such as the CDC-recognized Diabetes Prevention Program (DPP) to deliver coaching for diabetes and hypertension. Real-world deployments report clinically meaningful weight loss and blood-pressure reduction at 12 months, with a substantial fraction of participants living in rural areas and federally designated provider-shortage regions.
(Study summary)

These are not speculative pilots. They are examples where AI-enabled tools:

1. Deliver evidence-based interventions, and

2. Change workflows—who is reached, how often, and with what follow-up.

2.2 Directing scarce human effort more fairly

AI can also help prioritize limited human resources in chronic disease care:

• Risk stratification. Models that combine claims, electronic health records (EHRs), and social-risk data can identify patients at high risk for hospitalization or complications. When trained and evaluated on safety-net populations, they can focus nurse and care-manager time on people who would otherwise be missed.

• Remote monitoring for older and high-risk adults. The U.S. Veterans Health Administration’s long-running home telehealth programs—combining home devices, centralized monitoring, and care-coordinator workflows—have shown reduced hospitalizations and improved patient satisfaction, particularly among rural and mobility-limited veterans.

• Time-critical triage (e.g., stroke). AI-supported imaging and workflow systems can shorten “door-to-needle” times for stroke. From an equity lens, the key question is whether these gains occur in minority-serving and resource-limited hospitals, not just at well-resourced academic centers.

The principle: if you must ration human attention—and in healthcare you always do—AI can help you do it using better signals than historical utilization, provided you correct for biased labels.

3. FemTech, AgeTech, and what I’ll call “EquityTech”

Two domains where AI is already reshaping equity discussions are women’s health and aging. A third, less clearly labeled domain is technology aimed explicitly at under-served and under-represented populations.

3.1 FemTech: AI and women’s health

“FemTech” is a catch-all for technology-enabled products focused on health conditions unique to, or disproportionately affecting, women—menstruation, fertility, pregnancy, menopause, and beyond. (HealthTech overview)

The starting point is not neutral. A 2024 Nature editorial describes how women’s health research has historically been underfunded and under-prioritized, calling for dedicated funding streams and transparency.
(Nature – Women’s health research)

AI is being applied to these gaps in several ways:

• Maternal-risk prediction. Machine-learning models using EHR and ultrasound data are being developed to predict postpartum hemorrhage, pre-eclampsia, and other severe maternal outcomes, with the goal of earlier detection and targeted monitoring in high-risk pregnancies.

• AI-enabled obstetric ultrasound. Philips, supported by the Bill & Melinda Gates Foundation, has developed AI-powered tools on handheld ultrasound platforms like Lumify to help frontline workers in low- and middle-income countries identify high-risk pregnancies and refer patients appropriately. (Philips – AI-enabled Ultrasound)

At the same time, the FemTech boom has raised concerns about uneven clinical validation, affordability, and privacy, particularly with menstrual and fertility tracking apps. Reporting in outlets like The Washington Post has highlighted data-sharing practices and weak safeguards that could expose highly sensitive health information. (Washington Post – Femtech apps)

So FemTech is both an opportunity to correct deep neglect and a space where AI can easily become a thin layer over consumerism if not held to scientific and ethical standards.

3.2 AgeTech: AI for aging in place

“AgeTech” refers to technologies that support older adults’ health, independence, and social connection.

A 2022 review by Rubeis and colleagues describes AI-based AgeTech as a promising way to support aging in place, maintain social participation, and mitigate caregiver shortages—while warning that autonomy, dignity, and social context must be central design considerations.
(AI and AgeTech review)

Examples include:

• Fall-risk prediction and prevention

• Medication management and adherence support

• Cognitive-health monitoring and early detection of decline

• Companionship and communication tools to reduce loneliness

Pilot programs—such as converting televisions into AI-enabled “companions” that offer reminders, simple interactions, and video calls—illustrate both the potential and the risks around surveillance, consent, and digital literacy.

From an equity standpoint, AgeTech has to address affordability and accessibility from day one; otherwise, it becomes a premium service for affluent, tech-comfortable seniors.

3.3 EquityTech: centering the margins

We have labels for FemTech and AgeTech. We do not have a widely used term for technology built first and foremost for low-income communities and under-represented minorities—the populations with the largest health gaps.

For this discussion, I’ll use the term EquityTech: AI-enabled tools and systems deliberately designed for populations at the margins of health systems—low-income, historically excluded, or structurally disadvantaged groups—with equity impact treated as a core success criterion.

Concrete examples could include:

• Mobile tools for community health workers that work offline, in multiple languages, on low-cost phones, with AI assisting risk assessment and referral.

• Predictive models trained and validated on Medicaid, public-hospital, or rural-health data, with subgroup performance reported as a primary metric, not buried in an appendix.

• Remote-monitoring programs that provide devices and connectivity and integrate escalation pathways with local clinics, rather than assuming patients already have smartphones and broadband.

EquityTech is not a marketing label. It is a design stance: start from the margins, design outward, and measure success by what happens to the gap.

4. How AI can deepen inequity

AI does not automatically improve equity. In poorly designed systems, it can entrench and even magnify existing gaps.

4.1 Biased data and labels

A widely cited paper by Obermeyer et al. analyzed a commercial algorithm used to allocate extra care-management services in U.S. health systems. The algorithm used healthcare cost as a proxy for need. Because Black patients historically receive less care per unit of illness than white patients, the model systematically labeled Black patients as “lower risk” than equally sick white patients. (Science – Dissecting racial bias in an algorithm)

The lesson is straightforward: if you train models on spending and utilization, you will reproduce spending and utilization, not underlying clinical need. Equity-oriented AI must:

• define labels carefully (need, not cost), and

• report performance by race, ethnicity, sex, age, and socioeconomic status, not just overall accuracy.

4.2 Biased instruments

Measurement bias feeds directly into AI.

Pulse oximeters overestimate oxygen saturation in many patients with darker skin. Sjoding et al. showed that this leads to occult hypoxemia—arterial hypoxemia not detected by pulse oximetry—being significantly more common in Black patients. (NEJM – Racial Bias in Pulse Oximetry)

Any AI system that depends heavily on oxygen saturation must be audited by skin tone or ethnicity and adjusted—or it will simply encode this bias at scale.

4.3 Biased language models

LLMs used in healthcare applications can reproduce and magnify existing inequities.

The 2025 systematic review of LLMs in clinical medicine notes demographic bias and inconsistent behavior as key concerns, alongside hallucinations and variable accuracy. (PubMed – LLMs in Clinical Medicine)

Other studies have shown that LLM-based systems may:

• recommend more intensive diagnostics (e.g., MRI) for otherwise identical synthetic patients labeled as higher income

• vary empathy and thoroughness based on race, gender, or insurance status descriptors

This is not an edge-case concern; it has to be part of pre-deployment evaluation if LLMs are used for patient interaction or decision support. (Reuters summary)

4.4 Digital exclusion

Telehealth, apps, and remote monitoring will only improve equity if they explicitly address the digital divide.

Equity gains from telehealth will need workflows for patients without devices, broadband, or digital literacy, and non-app modalities such as SMS, telephone, and community access points.

A 2024 systematic review of digital health tools in rural populations found that while digital interventions can improve outcomes, they often exclude those with the greatest barriers unless device support, training, and low-bandwidth options are built in from the start. (PMC – Digital tech in rural health)

5. A practical equity-oriented AI playbook

What does it look like to take equity seriously in AI for chronic disease?

The arguments here, combined with guidance from the U.S. National Institute of Standards and Technology (NIST) and the World Health Organization (WHO), suggests a pragmatic playbook.

5.1 Measure inequity explicitly

• Build dashboards that show the care cascade—screening, diagnosis, treatment, adherence, escalation—stratified by race, ethnicity, sex, age, insurance type, and geography.

• Track data completeness and quality by subgroup. A model that performs well only for patients with complete data is not an equity solution.

5.2 Change workflows, not just models

• Choose specific breakpoints (for example, postpartum hypertension, untreated insomnia, or uncontrolled diabetes) and design end-to-end pathways: detection, intervention (digital and human), escalation, and feedback.

• Include non-app channels—SMS, telephone, community health workers, kiosks—from the beginning. If a solution assumes new smartphones, broadband, and English fluency, it is not an equity intervention.

5.3 Collect the right additional data

The whitepaper proposes a “new data minimum set” for credible equity claims:

1. Patient-reported outcomes (PROs) – sleep quality, pain, mood, functional status

2. Structured social-risk and barrier data – transportation, caregiving, cost concerns, housing, language

3. Device metadata and accuracy by subgroup – device type, firmware, and known performance differences

Collecting this data is not optional if equity is a stated goal; it is how we detect who is helped and who is harmed.

5.4 Use established governance frameworks

You do not need to invent governance from scratch:

• The NIST AI Risk Management Framework (AI RMF 1.0) offers a structured approach—govern, map, measure, manage—for managing AI risk, explicitly including fairness and safety. (NIST AI RMF)

• WHO’s Ethics and Governance of AI for Health guidance emphasizes human rights, accountability, transparency, and equity as core design principles. (WHO guidance)

• Civil-rights organizations such as the NAACP are now publishing “equity-first” expectations for health AI, calling for bias audits, transparency reports, and meaningful community engagement across the AI lifecycle. (Reuters – NAACP and AI in medicine)

If an AI initiative cannot be mapped onto frameworks like these, it is unlikely to deliver equitable outcomes.

5.5 Treat equity evidence as a product requirement

The chronic-care whitepaper notes that healthcare organizations and payers increasingly ask vendors for:

• Subgroup performance metrics

• Processes for monitoring drift and handling incidents

• Evidence that products work in real-world, diverse populations

Vendors that can supply this are starting to win contracts. Equity is slowly shifting from a “nice to have” to a competitive differentiator.

6. Will AI close health gaps?

AI will not, by itself, fix health inequity. It is not neutral. It amplifies the systems and values we embed into it.

But if we:

• Acknowledge current inequities with data rather than slogans

• Design AI around specific breaks in the care cascade

• Collect and use the additional data equity work requires

• Evaluate performance and safety by subgroup, not just overall

• Embed all of this within credible governance frameworks

…then AI can become a meaningful component of a broader health-equity strategy.

The three questions I now ask of any health-AI initiative are:

1. Which inequity is this system explicitly trying to reduce?

2. How will we know, for each subgroup, whether it succeeded or failed?

3. What safeguards are in place if it performs worst for the people we most want to help?

If we cannot answer those concretely, we are not doing equity work; we are running experiments on vulnerable populations.

In future Electric Sheep posts, I plan to go deeper on real-world implementations that take an “EquityTech-first” approach

For now, the simplest honest statement is:

AI is an amplifier. Whether it narrows or widens health gaps depends entirely on whether we treat equity as a design, deployment, and governance requirement—not an afterthought.

By the time I’m old enough for the senior menu, my android co-author will almost certainly have followed me home.

Right now, that “android” is just a large language model in the cloud that helps me draft essays like this one. We argue about structure; it suggests lines it thinks are profound; I cut most of them. It’s a brain in a box.

Fast-forward a couple of decades, and that same class of system is likely to be:

• listening for changes in my voice that hint at cognitive decline,

• watching my gait through a smartwatch and a smart hallway floor,

• negotiating with my cardiologist’s AI about my meds, and

• hiding inside practical hardware: a walker, a bed, maybe a robot that lifts me out of a chair without blowing out a caregiver’s back.

In Blood, Breath and Bits I argued that once you treat everything as “just another signal”—sleep, pulse, breathing, steps, clinical events—you can train models that see patterns humans miss. That’s already happening in sleep and chronic disease.

Aging is where that logic meets gravity.

If I put my physicist hat back on for a moment, the problem decomposes pretty cleanly. For most of us, old age fails along three coupled axes:

1. Mobility – whether you can move safely and independently.

2. Cognition – whether you can remember, plan, decide.

3. Chronic load – how many chronic conditions you’re carrying, and how volatile they are.

This essay is the sequel:

Bodies, Bolts, and Bits – how our aging bodies, mechanical helpers, and multimodal AI will have to work together if we want old age to be more than a slow-motion system failure.

Think of it as Blood, Breath and Bits… extended universe edition.

1. The Aging Baseline: Three Things That Give Way

Let’s start with the biology and the math. (I did train as a physicist; I can’t help myself.)

Demographically, we know the curve. The Population Reference Bureau estimates that Americans 65+ will grow from 58 million in 2022 to 82 million by 2050, with their share of the population rising from 17% to 23%. (PRB) That’s not a subtle trend; that’s a redesign-the-health-system signal.

Functionally, those extra years are complicated.

1.1 Mobility: not just “falls,” but how you move

National surveys and narrative reviews converge on the same picture: roughly 35–40% of people around age 70 report mobility limitations—difficulty walking a quarter mile or climbing ten steps—and the majority of those over 85 do. (PMC) Mobility limitations are consistently associated with increased fall risk, hospitalization, disability, lower quality of life, and higher mortality. (Frontiers)

Falls are the dramatic end point the CDC puts in headlines—over 38,000 deaths and nearly 3 million emergency-department visits for older adults in 2021 alone. (CDC) Underneath that is a quieter story: slowing, shrinking, and hesitating long before anyone hits the floor.

1.2 Cognition: the slow fuzzing of the edges

On the cognitive side, dementia is doing its own demographic climb.

The Alzheimer’s Association’s Facts and Figures report estimates that about 1 in 9 people 65 and older (≈11%) in the U.S. lives with Alzheimer’s dementia, with prevalence rising sharply with age. (Alzheimer’s Association) Global reviews put age-standardized dementia prevalence in people 60+ in the 5–7% range, with numbers projected to more than double by 2050. (PubMed)

That’s the careful, peer-reviewed version.

The Daily Show version is: at some point, you’ll start to wonder if the name you just forgot is normal aging or the beginning of something with a capital A.

1.3 Chronic load: the stack trace of diseases

Then there’s the chronic-disease pile-up.

A CDC-backed analysis of U.S. adults 65+ found that more than 50% of older adults and about 70% of Medicare beneficiaries have multimorbidity—two or more chronic conditions. (CDC) A recent global review in eClinicalMedicine suggests that in many high-income settings, multimorbidity prevalence in older adults often exceeds 60–70%, depending on how conditions are counted. (The Lancet)

This stack—hypertension, diabetes, heart disease, arthritis, chronic lung disease, depression—is strongly associated with:

• disability and functional decline, (CDC)

• higher hospitalization and readmission risk, (CDC)

• and lower quality of life and life expectancy as each condition is added. (PMC)

So my future self is not a clean “use case” in a product doc. I like to use the analogy of a three-legged stool: mobility, cognition, chronic load. Kick out any leg far enough and independence goes with it.

That’s the reality our android helpers will be dropped into.

2. Bodies and Bolts: Mobility as the First Pillar

Mobility is the first pillar because it’s the one you feel when you get up in the morning.

2.1 Measuring how we move (without turning everyone into an athlete)

A lot of the “robots for aging” don’t look like robots at all. They look like:

• a wristband,

• a patch,

• a sensor in the floor.

A 2021 systematic review in Journal of NeuroEngineering and Rehabilitation looked at how wearable sensors—accelerometers on the wrist, trunk, or ankles—have been used to assess frailty in older adults. It concluded that sensor-based metrics of gait, posture, and sit-to-stand performance provide objective and clinically meaningful frailty measures that align with conventional scales. (Springer Nature)

A 2025 review in Archives of Gerontology and Geriatrics and related work in Sensors and JMIR Aging extend this picture: simple, affordable wearables and activity trackers can screen for frailty, monitor mobility decline, and help prevent hospital-acquired disability in older adults. (ScienceDirect)

On the environment side, “smart” floors, inertial sensors, and in-home gait systems are being used to capture gait speed and stride variability during everyday life. Narrative reviews of mobility in older adults note that such objective measures correlate with frailty, falls, hospitalization, and mortality. (PMC)

Underneath the jargon, something simple is happening:

For a growing number of older adults, walking is turning into data that can help them.

2.2 Bolts that actually help, not just watch

Then you get devices that do more than observe.

Robotic hip exoskeletons, such as those tested in older adults and rehab settings, have been shown to improve walking efficiency and physical function and to reduce the metabolic cost of walking and stair climbing when compared with unassisted exercise. (JMIR Mental Health)

In Japan, RIKEN’s ROBEAR is an experimental nursing-care robot designed specifically to lift patients from bed to wheelchair and help them stand, providing “powerful yet gentle” assistance to offset caregiver strain in an aging society. (RIKEN)

These are not humanoids delivering monologues. They’re bolts arranged to extend mobility:

• by making walking less exhausting,

• by taking over dangerous transfers,

• by giving therapists extra leverage in rehab.

2.3 Bits: making sense of mobility over time

The “bits” piece is what I’ve spent much of my career working on: models that turn streams into signals.

In mobility, AI-based models increasingly combine:

• sensor-derived gait parameters,

• prior fall history,

• medication lists and comorbidities,

• and clinic-based test results,

to predict outcomes like falls, hospitalization, and disability. A 2022 systematic review in Sensors found that wearable sensor–based fall-risk assessment models can provide accurate, affordable surrogates for traditional fall-risk tools in community-dwelling older adults. (MDPI)

Reviews of mobility in older adults emphasize that declining gait speed and rising mobility limitations are strong independent predictors of later disability and death. (PMC)

Those models are the “bit” part of the mobility loop.

The LLM layer—my android co-author’s cousins—will be the thing that sits on top and says:

“Over the past three months, your walking speed has dropped and your balance has gotten shakier. People in your situation who do X, Y, Z—balance training, a medication review, a rail here and there—tend to stay independent longer. Want to talk about it?”

That’s mobility as a closed loop: bodies you can measure, bolts that can help, bits that can see patterns, and an android interpreter in the middle.

3. Bits and Speech: Cognitive Aging as the Second Pillar

The second leg of the stool is cognitive. This is where my internal science-fiction reader perks up.

3.1 Speech as a subtle sensor

A 2025 systematic review and meta-analysis in Age and Ageing looked at speech-based biomarkers for mild cognitive impairment (MCI) and dementia. It concluded that speech measures—lexical diversity, fluency, acoustic features—have good diagnostic utility for distinguishing MCI from cognitively unimpaired status when combined with clinical data. (OUP Academic)

Even more concretely, a 2024 study from the Framingham Heart Study applied automatic speech recognition and language models to recorded neuropsychological exams. The AI pipeline predicted which individuals with MCI would progress to Alzheimer’s within six years with about 78% accuracy, using only speech features and basic demographics. (PubMed)

Taken together, this line of work says:

• how quickly I find words,

• how often I lose the thread,

• how simple or complex my sentences are,

can give an AI system a statistically meaningful early warning that my brain is changing—well before a standard clinic visit would flag it. (Alzheimer’s Prevention Journal)

3.2 LLMs as cognitive scaffolding

Meanwhile, large language models are getting uncomfortably good at speaking medicine.

I talked about this in my inaugural post: A 2023 cross-sectional study in JAMA Internal Medicine compared real physicians’ answers to patient questions with ChatGPT’s answers. A panel of licensed clinicians preferred the chatbot’s responses nearly 4-to-1 and rated them significantly higher for both quality and empathy. (JAMA Network)

A 2024 systematic review in Journal of Medical Internet Research examined LLM-based systems across domains and concluded that they can exhibit elements of cognitive empathy—recognizing emotions and producing emotionally supportive responses—while also warning about safety and evaluation gaps. (JMIR)

Now combine that with speech:

• A speech model tracks changes in my language over months.

• An LLM sits on top and says, in so many words:

“Compared with a year ago, you’re taking longer to find words and losing your train of thought more often. People with similar patterns sometimes benefit from a memory assessment. Here’s what that involves—shall we schedule it?”

This is my android co-author turning into my cognitive scaffolding—not to replace my brain, but to notice when it might need help and explain that in ways that are less terrifying than late-night Googling or ChatGPTing.

Behind the scenes, of course, blood-based biomarkers for Alzheimer’s are advancing fast too—Nature Medicine recently reported highly accurate tau-based blood tests that correlate strongly with tau tangles and cognitive decline. (Nature) But day-to-day, it’s the voice and language layer that’s likely to be my main interface to cognitive AI.

4. The Chronic Stack: Multimorbidity as the Third Pillar

The third pillar is the least cinematic and probably the most important: the slow accumulation of chronic conditions.

4.1 Multimorbidity as the default setting

Systematic reviews across multiple countries tell us:

• In community-dwelling older adults, multimorbidity (two or more chronic conditions) often has median prevalence around 60–70%, with some cohorts reporting even higher rates in 75+ age groups. (PMC)

• In U.S. data, more than 50% of older adults and about 70% of Medicare

• beneficiaries have multimorbidity. (CDC)

Multimorbidity is strongly associated with:

• higher disability and functional limitations, (PMC)

• increased risk of new disabilities over time, (ScienceDirect)

• and a steeper decline in physical function and life expectancy as each new condition is added. (PMC)

This is the chronic stack: cardiovascular disease + diabetes + COPD + arthritis + kidney disease + depression, which rarely read each other’s guidelines.

4.2 Multimodal models for a multimorbid world

Here, multimodal AI does two important jobs.

First, it sees the whole stack.

Recent reviews of multimodal machine learning in healthcare—in npj Digital Medicine, Information Fusion, and Information—report that models which combine structured EHR data (diagnoses, labs, meds), imaging, time-series signals, and clinical text tend to outperform single-source models on tasks like diagnosis, prognosis, and risk stratification. (Nature)

That’s essentially Blood, Breath and Bits: instead of one model per disease, you get a model for this actual person’s trajectory, given everything that’s going on.

Second, it runs the loop.

We already have a mature example in older adults: hybrid closed-loop insulin delivery.

A 2022 randomized crossover trial in The Lancet Diabetes & Endocrinology compared hybrid closed-loop insulin delivery with sensor-augmented pump therapy in older adults with long-standing type 1 diabetes. The hybrid closed-loop system was safe and achieved significantly better time-in-range glucose without increasing hypoglycemia risk. (PubMed)

That’s what a working feedback loop looks like:

• bodies: glucose dynamics,

• bolts: continuous glucose monitor + pump,

• bits: control algorithm.

Now imagine similar patterns for:

• heart failure (weight, blood pressure, heart rate, symptoms),

• COPD (oxygen saturation, inhaler use, activity),

• hypertension (home BP, meds, side-effect reports).

Multimodal models forecast who is drifting into trouble; control systems propose or execute tweaks; LLMs explain why we’re adjusting that diuretic or beta blocker and what to watch for.

It’s messy, but it’s better than treating each chronic condition like it lives in a separate body.

4.3 Polypharmacy and mood

Then there’s the pharmacologic and psychological tangle.

Studies in geriatric populations routinely report polypharmacy (≥5 medications) in large majorities of multimorbid older adults, with clear links to adverse drug events and hospitalization risk. (Acr Journals) Depression, for its part, is highly comorbid with cardiovascular disease, diabetes, arthritis, and others, and independently predicts worse disability and quality of life. (ScienceDirect)

Here, LLM-driven systems and traditional ML have obvious jobs:

• cross-checking medication lists for high-risk combinations,

• surfacing deprescribing opportunities,

• and translating trade-offs into human language, so the person actually taking the medications has a say in their treatment.

5. When Bodies, Bolts, and Bits Sync Up

When you put all of this together—mobility, cognition, chronic load; sensors, robots, models, LLMs—you don’t get one magical android.

You get loops.

5.1 Mobility loops

• Bodies: my gait slows, my balance gets wobbly; stairs become a negotiation. (PMC)

• Bolts: a wearable and smart floor quietly track how I move; a robot walker or exoskeleton arm shows up in PT. (Springer Nature)

• Bits: models turn that into a mobility-risk profile; an LLM-based agent explains what’s happening and walks me through options—exercise, med review, assistive devices, home modifications. (MDPI)

The key difference from today is not that we care about mobility—that’s been true since humans invented chairs. It’s that we can see the decline early, continuously, and in context, and couple that to targeted actions.

5.2 Cognitive loops

• Bodies: my speech, habits, and digital behavior change in subtle ways.

• Bolts: a tabletop companion, smart speaker, or tablet collects enough audio (with my consent) for longitudinal analysis. Devices like PARO—the robotic seal used in dementia care—already show that carefully designed social robots can reduce agitation and improve mood in long-term care residents. (ScienceDirect)

• Bits: speech models detect patterns associated with early cognitive change; an LLM translates that into a conversation about assessment and planning instead of a cryptic score in an EHR. (OUP Academic)

If we’re thoughtful, that loop buys families years of planning time, not months of crisis.

5.3 Chronic-load loops

• Bodies: my conditions interact—diabetes worsens vascular disease; arthritis limits exercise; depression undercuts adherence. (Springer Nature)

• Bolts: monitors, inhalers, pumps, CPAP machines, blood-pressure cuffs all quietly emit data and implement therapy.

• Bits: multimodal models anticipate exacerbations; control algorithms fine-tune therapies (as in closed-loop insulin); an LLM explains the trade-offs so I’m not just passively swallowing pills and hoping for the best. (PubMed)

Across all of these, the android helper isn’t a humanoid standing in the corner.

It’s the coordination layer:

• the thing that sees across mobility, cognition, and chronic load,

• the thing that speaks human language in an empathetic manner,

• and the thing that can advocate for my preferences—whether I’m more afraid of falling, forgetting, or losing independence.

6. The Constraints We Can’t Code Around

If this all sounds a bit too much like too good to be true, sadly history and the literature gives us plenty of pause.

6.1 Autonomy vs “for your own good”

Aging is full of moments where safety and independence point in opposite directions:

• Do I keep driving or hand over the keys?

• Do I stay alone at home or move in with family?

• Do I accept a fall-detection camera in the bathroom?

Most AI papers focus on ROC curves and calibration plots. Very few deal with what it means to hand a 78-year-old a model-driven recommendation that says, in effect, “Stop doing this thing that makes you feel like yourself.”

Explainable-AI work in medicine and radiology emphasizes that clinicians need interpretable, context-rich models to maintain trust and accountability as AI systems interpret multimodal data. (PMC)

For android helpers in aging, I’d upgrade that to:

If a system can’t explain itself clearly to the person it’s guiding—in their language, with trade-offs they understand—it has no business reorganizing their lives.

LLMs give us the first credible way to do explanation at scale. Whether we use them that way is a governance choice, not a technical inevitability.

6.2 Inequity and health data poverty

Multimorbidity, mobility limits, and cognitive decline do not hit every community equally.

• A 2024 analysis of U.S. older adults found that the digital divide—lack of internet access and skills—remains persistent and is associated with worse self-rated health, greater functional limitations, multimorbidity, cognitive impairment, and depressive symptoms. (PMC)

• The U.S. Surgeon General’s 2023 advisory on loneliness highlights that lacking social connection increases the risk of premature death as much as smoking up to 15 cigarettes a day, and is associated with heart disease, stroke, dementia, depression, and anxiety. (HHS.gov)

On the AI side, The Lancet Digital Health Viewpoint on health data poverty argues that under-represented groups in health data risk missing out on benefits and being actively harmed by biased models if we don’t address gaps in who’s measured and how. (The Lancet)

If my android helper works wonderfully for people like me—plugged in, insured, English-speaking, middle class or upper, comfortable with tech—but misfires for low-income seniors or those outside the dominant datasets, that’s not an unfortunate side effect.

That’s building inequity into the operating system of aging.

Fixing it is partly technical (better datasets, bias audits, robust validation) and partly about power and policy (who gets devices, connectivity, training, and support).

6.3 Trust and emotional fit

Finally, there’s the human question: will anyone actually use this stuff?

Studies of older adults and digital health consistently list barriers like fear of “breaking” technology, low confidence, and privacy concerns, even when devices are clinically useful. (PMC)

At the same time, controlled experiments show LLM-based chatbots generating health explanations that human raters find more empathetic and higher quality than physician responses in online forums. (JAMA Network)

So we have the strange situation where:

• the back-end intelligence is getting kinder and more capable,

• while the front-end experience is often clumsy, infantilizing, or opaque.

“Android helper” is not just a technical system; it’s a relationship that has to feel respectful, not patronizing.

7. The Version of Old Age I’d Like to Meet

As I write this, I’m in that middle zone: old enough that these scenarios are starting to feel personal, young enough that I can still pretend they’re about “future me.”

My android co-author is still disembodied. It doesn’t know how fast I walk, how often I get up at night, or whether I saw another human being in person this week.

But if you look at where the science is pointing:

• Aging bodies are throwing off more signals than ever—from wearables, homes, and routine clinical care. (Springer Nature)

• Multimodal models are getting better at turning those signals into useful predictions across diseases and systems. (Nature)

• LLMs are getting better at knitting across those predictions and explaining them in language humans tolerate, and sometimes even like. (JAMA Network)

• Physical robots and smart environments are slowly learning to lift, steady, and comfort older adults without injuring them or their caregivers. (RIKEN)

That convergence—bodies, bolts, bits—is not a speculative sci-fi bet. It’s a conservative extrapolation of what’s already in PubMed, Nature journals, and the usual suspects.

My personal trajectory in this story is shifting from builder to observer to, eventually, subject.

So the questions I care about now are less “Can we do this?” and more:

• Will my android helper be there primarily to keep me alive, or to help me live the way I want for as long as possible?

• Will it be something only a thin slice of older adults can access, or will we treat it as infrastructure for aging in place?

• When it nudges me—about my walking, my memory, my meds—will it feel like a boss, a bureaucrat, or a partner?

Because at some point, if the actuaries and the universe cooperate, it won’t be a thought experiment.

It’ll be me, in an actual hallway, maybe holding onto a robot’s arm, while a web of models updates my mobility risk and drafts a note to my doctor. It’ll be my voice some system is parsing for signs of cognitive drift. It’ll be my chronic stack that an LLM is explaining back to me:

“Here’s what’s going on. Here are the options. What do you want to do?”

When that day comes, I’d like to recognize the system around me—not as something that just happened to me, but as something my generation of builders and leaders chose, deliberately, while we still had both feet under us.

Good science. Decent bolts. And a little respect for the humans in the loop.

In the second, I zoomed out to the systems level: blood, breath, and bits—and what happens when medicine finally sees the whole human at once.

This one is a little less romantic.

It’s about what I’ve had to learn, often by getting it wrong, about how AI actually ships in regulated healthcare—and why the same playbook is increasingly essential for any large organization that wants more AI value and less AI theater.

Over the years I’ve helped build my share of “innovation hubs,” skunkworks “AI labs,” and cool POCs for show-and-tell. The hope was: if we build something impressive enough, the rest of the organization will just get it.

Well, they didn’t.

Those failures were… educational. They’re a big part of why I now think in terms of operating systems, engines and interfaces, and organization-wide alignment, not just cool models.

My android co-author would like to claim it always knew this. It did not exist yet.

1. Pilots are cheap. Products are not.

If you sit in enough Board and ELT meetings, you start to hear the same line:

“We’ve got lots of GenAI pilots running.”

It sounds like progress. It usually isn’t.

Recent research from IDC and Lenovo, summarized in CIO and elsewhere, found that 88% of AI proofs of concept never make it to widescale deployment—for every 33 AI POCs, only four graduate to production. (cio.com)

A 2025 MIT study, The GenAI Divide: State of AI in Business 2025, reported that only about 5% of generative AI pilots deliver measurable P&L impact; 95% fail to move the needle, mostly because of poor integration, change management, and system design, not model capability. (MIT NANDA / State of AI in Business 2025, press coverage)

Healthcare starts from an even harder baseline:

• As of May 2024, the FDA had authorized 882 AI/ML-enabled medical devices across 510(k), De Novo, and PMA pathways. (U.S. Food and Drug Administration, FDA roundup, Regulatory Focus review)

• The EU AI Act now classifies most medical AI as high-risk, with explicit requirements on risk management, data governance, transparency, human oversight, and post-market monitoring. (Artificial Intelligence Act, High-level AI Act summary, EU digital strategy page)

So when I hear “we’re experimenting aggressively with GenAI,” what I actually hear is:

“We haven’t yet decided if AI is a core capability or a marketing campaign.”

Across finance, mobility, and now healthcare, I keep seeing the same pattern:

• AI experiments are easy.

• Revenue-driving AI products – particularly - in regulated environments - are hard.

Most of my own “innovation hub” adventures failed because I did not fully appreciate the latter.

2. A lab builds demos. An operating system ships products.

Most corporate AI journeys begin with a lab:

• a small, talented team

• some GPUs and notebooks

• a mandate to “explore use cases”

I’ve helped build those labs. They’re fun. They’re necessary. And they are very good at generating proof-of-concept gravity: demos, slideware, conference talks.

What they are not good at—on their own—is creating repeatable, governed, revenue-driving products.

If you stop at the lab, you get AI as artisanal craft:

• every model is bespoke

• every deployment needs heroic efforts

• standards are not followed

Six or twelve months later, you have an impressive graveyard of POCs…and very little that the manufacturing line, call center, clinic floor, or sales team can actually use.

What you actually need—especially in regulated healthcare—is an AiOS, an AI operating system.

By that I mean:

• A data backbone that can safely and reliably feed models with the right slices of EHR, device, and behavioral data.

• Boring, repeatable MLOps: model registries, CI/CD, deployment patterns, monitoring, rollback.

• Model governance as a living catalog: for each model, who owns it, what it’s for, how it was validated, where it runs, how it’s performing now. (U.S. Food and Drug Administration)

• A defined path from “interesting prototype” to “production feature” with product, clinical/operations, legal, and regulatory in the loop.

• Lifecycle ownership—someone whose job is to worry about that model after launch.

At Resmed, starting way back in 2020, we built on top of AWS infrastructure what we dubbed then as the Intelligent Health Signals (IHS) platform: an automated ML platform on Amazon SageMaker that greatly expanded AI/ML capabilities and streamlined model development and deployment across sleep and respiratory care. (AWS case overview). (Btw, with additional and much expanded GenAI services, the platform is now called ResAI.)

Under that simple description sits the real work: shared datasets, standardized pipelines, governance hooks, deployment patterns—an operating system that many different products and teams can stand on.

My earlier labs produced some beautiful prototypes. The operating-system approach produced things we could ship, audit, and scale.

That’s the first big lesson that generalizes far beyond healthcare:

If all you have is a lab, you will get pilots.
If you want products, you need an operating system.

Every large enterprise, be it a healthcare delivery org, a bank, an insurer, an energy or D2C company that is serious about AI is slowly rediscovering this, usually the hard way.

3. ML is the engine. GenAI is the interface.

In Electric Sheep #2, I wrote about foundation models built on sensor data and the rise of personal health agents—systems that combine continuous signals with language models to support people over time.

In practice, I’ve learned to think about the stack in two layers:

• The ML/DL layer is the engine. It predicts, scores, segments, recommends. These are the models you can, in principle, defend to the FDA, a regulator, or a clinical steering committee.

• The GenAI layer is the interface and conductor. It explains, orchestrates, converses, nudges. It sits between humans and the engine.

At Resmed, building products like myAir—our digital sleep companion for people on CPAP—started firmly at the engine level:

• models to predict therapy adherence

• models to stratify patients by risk and likely response

• models to recommend personalized comfort settings, trained on more than 100 million nights of real-world, de-identified sleep data—what became Personalized Therapy Comfort Settings (PTCS), now marketed as Smart Comfort. (ResMed Inc.)

Smart Comfort is the first FDA-cleared AI-enabled digital medical device that recommends personalized comfort settings to help people with obstructive sleep apnea start and stay on CPAP therapy.(Resmed press release, trade coverage)

On top of that engine, we layered:

• myAir’s personalized scores, insights, and coaching. A PwC-authored white paper and subsequent Resmed summaries show that myAir users, on average, use their device 46 minutes more per night and have higher therapy adherence than comparable patients without myAir (PwC myAir report, Resmed myAir adherence page)

• Dawn, Resmed’s generative-AI digital concierge, which provides instant, 24/7 answers about sleep health, Resmed products, and CPAP therapy support using GenAI.(Resmed press release on Dawn)

In other words:

• The engine optimizes for accuracy, calibration, robustness, and auditability.

• The interface optimizes for clarity, engagement, orchestration, and empathy.

Earlier in my career, I frequently over-indexed on one or the other: a gorgeous engine no one interacted with, or a delightful chatbot sitting on top of…not much. Healthcare is a very efficient teacher of humility.

This pattern generalizes:

• In finance, credit and fraud models are the engine; GenAI copilots for bankers and customers are the interface.

• In insurance, underwriting and pricing models are the engine; GenAI assistants for brokers and policyholders are the interface.

• In operations-heavy businesses, forecasting and optimization are the engine; GenAI agents for planners and frontline staff are the interface.

If your GenAI layer doesn’t sit on top of a serious engine, you’ve built a very charming steering wheel attached to nothing.

If you have a strong engine with no human-centric interface, you’ve parked a silent race car in a locked garage.

4. AI is changing how we work inside the walls

So far I’ve been tough on pilots and hype. Let me admit the other side.

AI—especially GenAI and agents—is already changing how we work internally:

• Engineers use AI pair-programmers to write, refactor, and test code faster (including for ML platforms and regulated software).

• Product managers and clinicians use AI to summarize evidence, draft specs, and enumerate edge cases before they brief a team.

• Ops and support teams use copilots to triage tickets, generate responses, and surface relevant policies or SOPs.

I feel this directly. Every Electric Sheep post, including this one, is co-written with my android:

• I bring scars, context, and narrative.

• It drafts, proposes structure, and fetches references.

• I argue with it until the result sounds like me, not like a panel discussion at an AI conference.

Inside large organizations, the same thing is happening to software factories:

• CI/CD and observability are being wrapped with AI agents that suggest tests, flag anomalies, and propose rollbacks.

• Documentation is evolving from static PDFs to conversations with systems that “know” your codebase and your data.

The MIT GenAI Divide report notes that many of the 5% of successful enterprise pilots are exactly these kinds of embedded copilots in specific workflows.

These internal gains matter. They compound. They often pay for themselves.

But they don’t change one stubborn fact:

Internal efficiency is necessary, but not sufficient.
At some point, AI has to show up in the P&L and in what patients or customers actually experience.

5. The real scoreboard is clinical and commercial, not just internal

Across sectors, we now have enough data to say what many practitioners have felt for years:

Most AI implementations today don’t move the needle for the business.

• IDC/Lenovo: 88% of AI POCs never reach widescale deployment; only 4 of 33 make it to production. (cio.com)

• Multiple analyses echo similar failure rates for “AI pilots” that never escape the slide deck. (Jeff Winter)

• MIT’s GenAI Divide report estimates that about 95% of GenAI initiatives show no measurable P&L impact, primarily because of weak integration and organizational readiness. (aigl.blog). While there us a lot of nuance to how this should be interpreted, the larger truth points to a real issue.

In healthcare, I’ve learned to look for a very short list of value levers. If a project doesn’t touch at least one, it’s probably theatre:

• Adherence and engagement – do more patients start and stay on therapy or in programs?

• Utilization and resupply – do the right patients get the right equipment, refills, and follow-up, with less waste?

• Workflow and staffing – do clinicians and staff get measurable time and cognitive load back?

• Revenue integrity – is documentation and reimbursement more accurate, efficient, and compliant?

myAir matters not because it’s pretty, but because data from a PwC report, a Resmed white paper, and subsequent summaries show that myAir patients use their devices longer per night and have higher adherence than those without it.
(PwC report PDF, Resmed myAir press release)

Smart Comfort matters not because it’s “AI-enabled,” but because it uses models trained on over 100 million nights of real sleep data to personalize comfort settings—directly addressing a major dropout point in CPAP therapy. (ResMed Inc.)

Dawn matters not because it’s GenAI, but because it provides instant, 24/7, personalized support at scale—helping patients navigate therapy and taking load off human teams.

Inside the company, AI changes how we work. Outside, it has to change how value is created and captured—clinically and commercially.

That’s as true for healthcare as it is for banking, logistics, energy, or SaaS.

6. Alignment from boardroom to manufacturing line

There’s one more lesson that keeps being reinforced:

You cannot build a serious AI capability as a side project tucked into a corner of the org chart.

In regulated healthcare, AI is now clearly a board-level topic:

• Risk, because bad AI can harm patients, reputations, and balance sheets.

• Opportunity, because good AI can reshape how care is delivered and how value flows.

The same is increasingly true in any AI-ambitious industry—whether your “patient” is a person, a portfolio, or a production line.

Getting this right means alignment from the Boardroom all the way to the manufacturing line, with every function in between—R&D, regulatory, quality, clinical, operations, supply chain, sales, marketing, finance, IT—pulling in the same direction.

In practice, that looks like:

• Board and ELT clarity

  • AI is on the agenda regularly, not just during hype spikes.

  • The Board understands the basic risk and opportunity profile using your products and markets, not generic AI slides.

  • There is a clearly accountable senior leader (CAIO, CDAO, etc.) responsible for the AI operating system: platform, guardrails, portfolio, and outcomes.

• Middle management and functional buy-in

  • VPs and directors know which parts of their world are “AI-first” candidates and which aren’t.

  • Functions see AI as part of their toolkit for hitting targets—whether that’s defect rates on a manufacturing line, readmission rates in a population, or margin in a segment—rather than something “the data team does over there.”

• Education and upskilling as part of the OS, not a side quest
  One of the big mistakes in my earlier “innovation hub” days was assuming that impressive demos would magically spread understanding. They don’t. What does help is treating education as infrastructure:

  • short, tailored sessions for the board and ELT on what AI can and cannot do in this company

  • role-specific training for:

    • clinicians and operators (how to interpret model outputs, when to override them),

    • manufacturing and supply-chain teams (how AI forecasts and anomaly detectors fit into daily routines),

    • sales and marketing (how to use AI safely in campaigns, forecasts, and customer interactions)

  • internal “AI roadshows” across sites and functions, where people can see and challenge real systems—not just see a keynote.

• Skilled practitioners embedded everywhere
  Over time, organizations that get good at this don’t centralize AI in one heroic team. They develop AI-fluent practitioners in every key function:

  • product managers who can frame ML problems,

  • clinicians who understand what a risk model really is,

  • ops leaders who can reason about trade-offs,

  • manufacturing engineers who can work with predictive maintenance and quality models.

The central AI/ML group still builds and stewards the operating system, but success looks like more and more of the work being done with that system by people in the line of business.

When this alignment clicks—from Boardroom to manufacturing line—three things become possible:

1. Massive impact. Because you’re aiming AI at the levers that actually move clinical and commercial outcomes.

2. Repeatable impact. Because you’re not reinventing governance and infrastructure for every new idea.

3. Fast impact. Even in healthcare AI—one of the slowest, most regulated arenas—you can move quickly when alignment, platform, and skills are in place.

It’s not glamorous work. But it’s the difference between “we have an innovation lab” and “we can actually move the needle with AI.”

7. A playbook that travels

Everything in this post has been grounded in regulated healthcare because that’s where I’ve been operating.

But the playbook itself travels:

• Operating system, not lab. You need shared data, MLOps, and governance anywhere the stakes are real.

• Engine + interface. Grounded ML for decisions; GenAI for explanation, orchestration, and behavior change.

• Internal efficiency + external economics. Use AI to change how you work and how you create and capture value.

• Alignment and education. From board/ELT to manufacturing line, everyone needs a mental model of what AI is doing and why.

• Practitioners everywhere. The more AI-fluent people you have in product, ops, clinical, manufacturing, legal, finance, and sales, the stronger the whole system becomes.

In regulated industries—healthcare, finance, insurance, energy, transportation—this is non-negotiable. Regulators are already treating AI systems as safety-critical artifacts with expectations around documentation, oversight, and post-market surveillance. (RAPS Brand Portal)

In unregulated domains, this same structure is how you quietly build Responsible and Ethical AI without turning it into a separate religion:

• An OS so you know where your models are, what data they see, and how they behave.

• A clear separation between engines you can audit and interfaces you can constrain.

• A habit of tying projects to value levers that matter to humans, not just dashboards.

• A culture where legal, security, UX, and domain experts are there from day one, not bolted on at the end.

This is the through-line back to Electric Sheep #1 and #2:

• In #1, I asked what happens when you write with an android.

• In #2, I asked what happens when medicine sees your whole life as data.

• In this one, I’m really asking:

What does it take to build organizations that deserve that power?

Part of my answer is this operating-system playbook, forged in a mix of successful launches and very public “learning opportunities” in labs and innovation hubs.

The rest, we’ll keep unpacking—android co-author and all—in future installments of Electric Sheep.

This is the follow-through.

Today I want to explore a simple question that turns out to be anything but:

What happens when medicine finally sees the whole human—blood, breath, sleep, movement, mood, history—all at once, and AI can actually make sense of it?

Across research labs and a growing number of products, three threads are starting to braid together:

• We’re collecting ubiquitous, multimodal, multi-scale data about our bodies and lives.

• We’re building foundation models based on sensor data, not just text and images.

• We’re prototyping personal health agents that sit on top of all this and talk to us in plain language.

This post is my attempt to sketch where I think that convergence is going, before zooming into specific topics—sleep, chronic disease, mental health, aging—in later installments.

From snapshots to streams

For most of modern medicine, you and I show up as a stack of snapshots:

• an annual lab panel

• a sleep study (once, maybe)

• a clinic note written in a hurry

• a medication list that may or may not be correct

It’s better than nothing, but it’s still a low-frame-rate movie of a very fast, very complex system.

Now compare that to what we could see:

• High-resolution physiology: brain activity, heart rhythms, respiratory flow, oxygen saturation, movement, heart rate variability, breathing sounds

• Omics: genomics, proteomics, metabolomics, microbiome

• Digital records: EHR data, imaging reports, prescriptions, clinical notes

• Everyday signals: sleep and activity from wearables, continuous glucose monitors, home blood pressure cuffs, smartphone sensors, voice and language patterns

Sleep is a good microcosm of this shift.

Thapa and colleagues recently introduced SleepFM, a multimodal sleep foundation model trained on more than 500,000 hours of high-quality sleep recordings from roughly 65,000 participants across multiple cohorts. (PubMed Central)

Instead of building one bespoke model per signal and per task, they train a single model on full polysomnography—brain, heart, and respiratory signals together—then reuse its embeddings for many downstream problems. Those embeddings don’t just help with things like sleep staging; they also improve prediction of 130 future diseases, with respectable discrimination for outcomes like heart failure, chronic kidney disease, dementia, stroke, and myocardial infarction.

In parallel, Fox et al. have built PFTSleep, a self-supervised transformer that ingests entire eight-hour sleep studies—brain, movement, cardiac, oxygen, and respiratory channels—and then classifies sleep stages for the whole night in one pass. In validation across several large cohorts, their approach improved sleep staging performance compared with existing methods. (OUP Academic)

In plainer English: we’re finally starting to train models on the full movie of a night’s sleep, not just isolated frames.

And this isn’t just happening in academic papers.

The U.S. FDA has cleared Apple’s Sleep Apnea Notification Feature on the Apple Watch as an over-the-counter device to assess risk of moderate to severe sleep apnea. Samsung’s Galaxy Watch has a cleared sleep apnea feature as well, via a separate regulatory pathway. (FDA Access Data)

Meanwhile, devices like the Oura Ring have been studied head-to-head against in-lab polysomnography and actigraphy; meta-analyses and validation studies show comparable accuracy for many commonly used sleep parameters. (PubMed Central)

So your watch, ring, band, or CPAP machine is no longer just a gadget. It’s quietly turning into a clinical-grade sensor—whether your health system has caught up or not.

From large language models to “large sensor models”

Large language models taught the world that if you give a system enough text, it starts to pick up reusable abstractions about language and the world.

We’re now seeing the same pattern with continuous sensor data.

SleepFM is a nice example of what I’d call a large sensor model:

• It takes in multimodal time-series (EEG, ECG, respiration).

• It learns a shared representation of the body’s dynamics during sleep.

• Those representations are reusable across tasks: staging, sleep-disordered breathing, and even longer-term disease prediction.

PFTSleep pushes in a similar direction: encode the entire night with a single transformer, then let simple heads do the task-specific work.

The conceptual shift is important:

Instead of “one model per metric per task,” we’re learning general physiologic embeddings and reusing them.

A parallel thread is emerging at the interface of sensors and language.

In Nature Medicine, Khasentino et al. introduced the Personal Health Large Language Model (PH-LLM), a version of Gemini fine-tuned for sleep and fitness using wearable data. (PubMed)

A few notable results:

• On multiple-choice exams, PH-LLM scored 79% vs 76% for human experts on sleep medicine questions and 88% vs 71% on fitness questions.

• In 857 real-world case studies, its long-form coaching responses were rated roughly on par with human specialists for fitness, and close to expert quality for sleep.

• It could predict self-reported sleep quality from multimodal wearable data, showing it actually learned from the numbers, not just guideline text.

If you look at their main schematic, the pipeline is straightforward:

1. Daily wearable metrics are encoded into compact features.

2. PH-LLM reasons over those features and text.

3. The output is a conversational explanation and coaching plan that looks more like something a human coach might write than a lab report.

SleepFM and PFTSleep are strong sensor backbones. PH-LLM shows that an LLM can sit on top and “speak sensor” fluently.

The last piece is orchestration.

Enter personal health agents

This is where personal health agents come in.

In The Anatomy of a Personal Health Agent, Heydari and colleagues (Google Research and collaborators) propose a multi-agent system that reasons over wearable data, simple health records, and user goals to deliver personalized health support. (arXiv)

They split the work across three collaborating agents:

• A Data Science Agent that analyzes your time-series and population statistics

• A Health Domain Expert Agent that integrates medical knowledge with your personal context

• A Health Coach Agent that turns those insights into behavior-change-oriented recommendations and tracks your progress over time

An orchestrator routes your question through these agents and stitches their outputs into a coherent conversation.

Crucially, they evaluated this system across 10 benchmark tasks, with over 7,000 human annotations and more than 1,100 hours of effort from health experts and end-users—probably the most comprehensive evaluation of a consumer-facing health agent to date.

Put this together with PH-LLM and SleepFM-style models and you get something that behaves less like a single chatbot and more like a small, always-on care team living in your phone.

The natural next question is: what does that actually look like in day-to-day life?

How this might show up in everyday care

I’ll resist the temptation to write “A Day in the Life of Future Badri” (I’ll save that for another post) and instead talk through a few realistic patterns that fall straight out of the current evidence.

1. Diagnostic “front doors” on devices you already own

The sleep world is already halfway there.

The FDA has cleared Apple’s Sleep Apnea Notification Feature as an over-the-counter tool to assess risk of moderate to severe sleep apnea on consumer watches, based on comparison against formal sleep testing. Samsung’s Galaxy Watch has its own cleared sleep apnea feature, classified through a de novo process. (FDA Access Data)

Meanwhile, large validation studies show that devices like Oura can approximate polysomnography and actigraphy for key sleep measures at a population level, and that ring-based systems can help detect sleep apnea when compared with PSG. (PubMed Central)

Combine that hardware with large sensor models trained on hundreds of thousands of hours of PSG, and you can sketch a near-term workflow:

• Your watch and ring continuously collect sleep and cardio-respiratory signals.

• A foundation model based on sensor data quietly screens for patterns consistent with sleep apnea, insomnia subtypes, periodic limb movements, circadian disruption, and more.

• A personal health agent explains the findings in plain language (“Here’s what we measured; here’s what we think it means; here’s how sure we are”) and suggests next steps—anything from reassurance, to a home sleep test, to a referral for in-lab polysomnography.

Nothing in that story requires sci-fi. All the technical ingredients exist today.

The same pattern extends well beyond sleep:

• In cardiovascular medicine, reviews in journals like Circulation Research and JACC describe how wearables plus machine learning can support arrhythmia detection, blood pressure monitoring, and risk stratification for cardiometabolic disease. (American Heart Association Journals)

• In chronic disease management more broadly, scoping reviews show wearables being used to monitor physical activity, symptoms, and vitals across conditions like diabetes, COPD, and heart failure—though with gaps in who gets access and which diseases are covered. (PubMed Central)

Once you treat these as just different “sensor channels,” sleep stops being an isolated use case and becomes the first in a long line of diagnostic front doors.

2. A longitudinal “risk radar” for chronic disease

One of the most interesting aspects of SleepFM is that its embeddings—learned from raw sleep data—could predict a wide range of future diseases more accurately than traditional hand-crafted features.

Generalizing the idea: if we combine

• sleep patterns

• activity and sedentary behavior

• cardiopulmonary signals

• labs, meds, and comorbidities

we can start to construct digital phenotypes that track how someone is actually doing over months and years, not just in a 12-minute visit.

Digital phenotyping work in mental health has already shown that passive signals—sleep, mobility, phone use, heart rate—can help monitor depression, bipolar disorder, and other conditions. (PubMed Central) Reviews in digital phenotyping and wearable-based chronic disease management make similar points for cardiometabolic risk and other long-term conditions, despite some concerns. (PubMed Central)

A personal health agent on top of that could do things like:

• Track trends, not just thresholds (“your sleep efficiency and activity have steadily worsened over six months, and in people like you, that often precedes a rise in A1c”).

• Draft a brief, structured note for your clinician summarizing what’s changed and why it might matter.

• Suggest next actions: a medication review, extra labs, a trial of CBT-I, a blood pressure check, or a referral.

This is where the convergence really matters. Sleep, movement, metabolic markers, and mood are intertwined. Treating them as separate specialty silos is a billing artifact, not a biological truth.

3. Everyday coaching that actually knows you

On the coaching side, we’re already well past one-size-fits-all sleep hygiene tips.

Take myAir, a Resmed patient engagement app for people on CPAP therapy. It gives users nightly scores, tracks trends, and uses interactive, tailored coaching to improve adherence—studies have found significantly higher CPAP usage among myAir users compared with those receiving standard care alone. (Chest Journal)

Resmed has also launched Dawn, a generative-AI–enabled sleep health concierge that uses large language models to provide personalized guidance and answer questions about sleep health and therapy. (ResMed Inc.)

If you zoom out from specific brands, the pattern is clear: AI-supported coaching is starting to sit between you and a full-blown clinical visit, especially for conditions that are heavily behavioral.

And we know that digital CBT-I works. Programs like Sleepio and other dCBT-I platforms have been tested in multiple randomized trials and real-world studies, showing substantial and durable improvements in insomnia symptoms and sleep efficiency. (Big Health) Recent reviews explicitly discuss how AI and large language models could make these interventions more adaptive and personalized over time. (ScienceDirect)

Put PH-LLM-style reasoning, PHA-style orchestration, and dCBT-I evidence together, and you can imagine:

• A sleep and stress coach that understands your personal patterns rather than generic averages

• Guidance that ties symptoms to concrete behaviors (“on nights you nap after 5 pm, both your sleep efficiency and next-day mood drop”)

• A clear hand-off to clinicians when the problem goes beyond what a coach or app should handle

Again, this isn’t just about sleep. Similar architectures are being explored for diabetes, hypertension, obesity, and other chronic conditions. Sleep is simply an especially rich—and underused—signal in that larger story.

Why I’m excited and cautious

As someone who’s been building applied AI systems for almost three decades now, I’ve seen my share of hype cycles.

This one feels different for a few reasons:

• The data really are bigger and richer. SleepFM isn’t trained on a few hundred overnight studies; it learns from over 500,000 hours of sleep data across roughly 65,000 people, spanning multiple cohorts and settings.

• We finally have foundation models based on sensor data. Instead of bespoke models for each metric, we can train once on massive multimodal time-series and reuse those representations across tasks.

• LLMs are starting to truly integrate sensors. PH-LLM is not “ChatGPT with a health prompt”; it’s explicitly trained to interpret daily wearable data and combine that with domain knowledge, performing at or above human experts on structured exams.

• People are taking agents seriously, not just models. The PHA work treats “helping someone live healthier” as a multi-agent, multi-task problem and evaluates it end-to-end with real user questions and extensive expert review.

At the same time, a few caution lights are blinking:

• Not everyone is wearing a $400 watch. Reviews of wearables in chronic disease management highlight clear disparities in who uses these devices and which conditions get attention. (i-jmr.org) If we’re not careful, we’ll build exquisitely personalized systems for the already well-served.

• Socioeconomic determinants don’t disappear because you added an app. Digital phenotyping can capture behavior and some context, but it doesn’t fix food deserts, unsafe housing, or precarious work. Recent work on digital phenotyping and mental health emphasizes exactly these limitations—even as it shows real promise. (PubMed Central)

• Regulation and accountability are still murky. When a personal health agent reassures you that something is fine—or nudges you toward a choice that turns out badly—where does responsibility sit? With the model developer, the clinician, the hospital, the app vendor, you? We’re only starting to wrestle with that.

• Human behavior is messy. Even perfectly calibrated insights don’t automatically translate into adherence or trust. That’s why the coaching side of PH-LLM and PHA—informed by behavioral science and evaluated with real people—matters as much as the ROC curves.

So no, I don’t think “AI will save healthcare.”

My view is more modest—and more interesting to me:

If we’re deliberate, this convergence can move us from episodic, siloed care to continuous, whole-person support—especially for chronic disease, sleep, mental health, and aging.

That’s worth some cautious excitement.

Where I’ll probably go next (in this Substack)

This post was intentionally broad—a map of the territory as I see it in late 2025.

Over the next few Electric Sheep posts, I expect to zoom in on threads like:

• Sleep as a longitudinal vital sign – how foundation models based on sensor data could connect night-to-night patterns with cardiometabolic and mental health trajectories.

• Personal health agents in practice – what it would actually take (technically, clinically, and legally) to deploy PH-LLM / PHA-style systems in the real world.

• Chronic conditions in combination – how an integrated approach might look for people (like me) living with overlapping issues such as sleep apnea, type 2 diabetes, and cardiovascular risk.

• Health equity and the social determinants of health – how to avoid creating a two-tier system where the well-insured with wearables get precision digital care and everyone else gets whatever’s left.

• Aging and brain health – what it means to forecast long-term cognitive risk from mundane signals like sleep and daily activity, and whether we’re ready for that.

• Failure modes and governance – bias, over-medicalization of normal life, “alert fatigue 2.0,” and what good guardrails might look like.

Some of the ideas I’m excited about today will quietly fail in pilots or peer review. Others will become boring infrastructure. The narrative will wander a bit, just like the real world does.

My plan is to track that evolution as both:

• someone who builds these systems, and

• someone who lives in a body those systems increasingly measure.

If that intersection of blood, breath, and bits intrigues you, stick around. The sheep are just getting started.

Subscribe now

Electric Sheep

On AI, android co-authors, and how it all began with sci-fi

When I was a teenager in Delhi, I met my first android.

Not in a lab. In a dog-eared paperback of Philip K. Dick’s Do Androids Dream of Electric Sheep? bought from a second-hand bookstall. That world—radioactive dust, artificial animals, almost-human androids—stuck in my head and quietly rewired it.

That book, and later the movie Blade Runner based on the book, did two things to me.

First, they nudged me toward physics. If futures could look that strange, I wanted to understand the laws underneath them. So I became a theoretical physicist.

Second, they made me deeply suspicious of neat stories about technology. In Dick’s world, humans become inhuman, androids become uncomfortably human, and empathy is supposedly the one test that separates us—until it doesn’t.

So when I decided to start this Substack, Electric Sheep felt inevitable.

It’s a nod to the sci-fi that shaped me, but it’s also literal: this is an open co-creation between me and my “android” co-author—an AI assistant. You’re reading both of us right now.

And this will be a running series on AI as I actually see it in the wild: in healthcare settings, in product roadmaps, in boardrooms, and increasingly, in our heads.

Why Electric Sheep?

Dick’s universe is full of convincing fakes: fake animals, fake memories, fake emotions. The uncomfortable question underneath is simple:

When our world fills with simulations, what does it mean to be human?

We’re now running that experiment in real time.

We’ve built systems that can mimic reasoning, style, even vulnerability. They write notes, generate images, “listen” to patients, and help people feel heard at 2 a.m. on the internet.

Sometimes, disturbingly, they seem more caring than the humans they’re augmenting.

A recent study led by UC San Diego researchers compared answers from physicians on a public medical forum with responses generated by an AI assistant. A panel of licensed healthcare professionals rated the AI’s answers as both higher quality and far more empathetic than the doctors’ responses; empathetic or very empathetic answers were several times more common from the chatbot. The work was published in JAMA Internal Medicine in 2023 (Ayers et al., 183(6):589–596, doi:10.1001/jamainternmed.2023.1838).

Read that again: clinicians preferred the chatbot’s bedside manner.

As someone who’s spent years building AI in healthcare, I find that both promising and deeply unsettling.

Promising, because if we can give burned-out clinicians a draft that is accurate and kind, maybe more patients get good answers faster—and more doctors get home for dinner.

Unsettling, because empathy reduced to a pattern we can mass-produce forces us to revisit what we think empathy actually is. If a statistically trained model can out-empathize a human doctor in text, what else about our “uniquely human” territory is suddenly negotiable?

This is exactly the kind of tension I want to sit with in Electric Sheep.

Meet my android co-author

Let me be explicit about the arrangement here.

I’ve been doing AI long enough to remember when “neural networks” still sounded fringe and “AI” was something only marketing people said out loud.

I started in theoretical physics, moved into applied machine learning in financial services, then co-founded a predictive analytics startup. I’ve led AI for a large mobility platform where every model decision played out across millions of rides a day. And most recently, I’ve been leading AI in a global medtech company focused on sleep and respiratory care.

In other words: I’m not writing about AI from the balcony, nor is it a newfound religion. I’m usually somewhere backstage, trying to keep the machinery from catching fire.

For most of that career, AI lived behind the curtain. Models made decisions, but the interface was a report, a dashboard, or a line in a transaction log.

Now the model is in the room.

It’s answering emails. Drafting policies. Summarizing medical notes. Helping me structure this essay.

So rather than pretend otherwise, I’m treating my AI assistant as a visible collaborator:

• I use it to draft, edit, and challenge my thinking.

• I keep the final editorial responsibility. If something sounds wrong or glib, that’s on me.

Part of the experiment of Electric Sheep is to show, transparently, what it looks like for a domain expert and a very capable, occasionally clueless model to work together.

I don’t think leaders—or citizens—can afford to talk about AI in the abstract anymore. We need to use it, feel its strengths and its blind spots, and then make adult decisions about where it belongs.

From physics to applied AI (and why I’m doing this now)

A bit of personal context.

Sci-fi got me dreaming about alternate futures.

Physics taught me the universe doesn’t care about dreams (although it is frequently foretold in those); it cares about equations. My PhD work was in theoretical physics, studying systems that are deterministic and chaotic at the same time. In hindsight, that was excellent training for two things: complex organizations and modern machine learning.

From there I moved into applied AI:

• Early neural networks for credit risk and forecasting.

• Startup life, building analytics products from scratch.

• Large-scale AI for mobility and marketplaces.

• And then healthcare and sleep—where your models don’t just move click-through rates, they influence whether someone wakes up tomorrow feeling like themselves.

And here’s the hopeful part.

We are, for the first time, at a genuine convergence point:

• Multidimensional, multimodal, ubiquitous data about our bodies and our lives—from medical devices, wearables, sensors, and electronic health records.

• Cutting-edge algorithms that can make sense of that data across time, space, and modality.

• Compute infrastructure that can actually run these models at the scale of real populations, not just research cohorts.

Chronic diseases that once felt “too complex” to do anything about beyond late-stage management are starting to look analytically tractable. AI is not a magic wand here, but it is uniquely suited to spotting patterns, trajectories, and early warning signs that humans simply can’t see unaided.

I’m starting Electric Sheep now because we’ve crossed a line.

We’re past “AI might be important someday.” We’re in the phase where the choices we make—about regulation, deployment, incentives, and culture—will lock in trajectories that are very hard to reverse, including how we fight (or fail to fight) chronic disease.

I want to write about that moment in real time, as a practicing AI leader, not as a historian looking backward.

And I want to write it for anyone who feels they’ve been handed a front-row seat to a transformation they don’t fully understand yet: executives, clinicians, builders, policy folks, and the simply curious.

How I’ll try to write here

A few ground rules for Electric Sheep:

1. No worship, no panic.
   I’ve seen AI do extraordinary things. I’ve also seen it overfit, hallucinate, and quietly fail in production while everyone was busy celebrating the demo. I’m not here to sell you salvation or doom.

2. Evidence over vibes.
   When I say “this works” or “this is risky,” I’ll lean on data, real deployments, or lived experience—not just conference slides.

3. Real-world impact first.
   I care about what AI does to people: patients trying to understand their lab results, product teams under pressure, overworked clinicians, young and older workers worried about their careers.

4. Write what I know.
   I’ll mostly stay in the domains I’ve actually lived in—data science, AI in real organizations, health and sleep, and the leadership and governance messiness around them. When I speculate, I’ll try to label it as such.

5. Permission to be messy.
   I’m not going to force this into a neat consulting-style framework. Real life doesn’t cooperate, and neither does real AI.

What I’ll probably write about (subject to chaos)

I have a rough sketch for where Electric Sheep might go. Emphasis on might.

Expect posts on things like:

• AI and sleep as a vital sign – how billions of nights of sleep data could reshape how we think about chronic disease, mental health, and aging.

• How AI actually ships inside regulated companies – governance, platforms, and the unglamorous plumbing required to get from “cool model” to “approved product in clinic.”

• Empathy at scale – what studies like the UCSD work cited previously on chatbot vs physician responses tell us, and what they absolutely don’t tell us, about care and clinical reality.

• Making chronic disease less inevitable – using ubiquitous, multimodal data plus modern algorithms and compute to detect, prevent, or slow disease instead of just documenting it.

• The costs we’d prefer not to think about – environmental, social, and psychological.

But I’m also leaving room for this to wander.

Like any good sci-fi novel—or life—the plot will zigzag. Some posts will start in sleep medicine and end in regulatory philosophy. Others will begin with a model architecture and end with a story about a patient, a doctor, or an AI engineer.

If you’re looking for a perfectly structured “AI in 10 easy lessons” curriculum, this is not that.

If you’re okay following a practitioner and his android through the fog, trying to make sense of what we’re building as we build it, then you’re exactly who I’m writing for.

One last question

Do androids dream of electric sheep? I don’t know what, if anything, today’s models “dream” of. The safest assumption is: nothing. They are staggeringly sophisticated pattern machines, not people.

But the more interesting question to me is this:

What kind of humans do we become when we live and work alongside these systems every day?

That’s what I want to explore here—honestly, occasionally skeptically, and with the help of an android who, according to at least one UCSD study, is not only smarter on average, but asounds kinder than many physicians we know.

If that tension intrigues you, welcome to Electric Sheep.