There is a version of HealthTech that I genuinely believe in. One where a patient in a rural village gets the same diagnostic accuracy as someone at a major urban hospital. Where a first-generation immigrant can access care in their language without a human intermediary. Where the quality of your healthcare is decoupled, for the first time, from your postcode. That version is technically possible. In many ways, it is closer than it has ever been.
The version we are actually building looks quite different. Today, the same AI tools meant to extend and democratise care are quietly encoding decades of structural inequality into their predictions. They are being trained on data that systematically underrepresents the people who need them most — and then deployed as if that underrepresentation does not matter. It does. It matters in the most concrete, consequential sense: in missed diagnoses, in withheld treatment, in patients sent home when they should have been admitted.
I want to talk about that gap — between what HealthTech promises and what it currently delivers — and about what it would actually take to close it.
The Data Was Never Neutral
The foundational problem is this: machine learning models in healthcare learn from historical data, and historical data reflects historical inequity. When a model is trained on who received care, how much was spent on them, and what outcomes they experienced, it is not learning medicine — it is learning the contours of a system that has, for generations, provided less care to Black, Brown, Indigenous, low-income, and rural patients. The model has no way to distinguish between "this patient is healthier" and "this patient was given fewer resources." It treats both as the same signal.
The landmark demonstration of this was a 2019 study published in Science by Obermeyer et al. Researchers analysed a widely deployed algorithm used by US hospitals to identify high-risk patients for care management programmes. The algorithm used past healthcare spending as a proxy for current health need — a design choice that appeared neutral and data-driven. It was not. Because Black patients had historically spent less on healthcare than equally sick white patients — a direct consequence of wealth disparities rooted in structural racism — the model systematically underestimated their health risk. The result was that Black patients had to be measurably sicker than white patients to be referred for the same level of care. The algorithm did not invent that disparity. It automated it at scale.
"Algorithms cannot be trusted to make safe and fair decisions about patient care when they are trained on data that already encodes who society decided was worth treating."
— Amber Nigam, Co-founder & CEO, basys.ai (STAT News, November 2024)
This is not an isolated incident. A systematic review published in the Journal of Racial and Ethnic Health Disparities (2026) examined 30 studies on AI-driven racial disparities in healthcare and found a consistent pattern: biased training data, homogenous development teams, and the uncritical use of race as a proxy variable were the primary drivers. The same review found that Black and Hispanic patients were the populations most frequently and most severely disadvantaged. These are not edge cases. They are the rule.
The Device on Your Finger
The bias in HealthTech is not only algorithmic. It is also physical. During the COVID-19 pandemic, an uncomfortable truth surfaced about pulse oximeters — the small devices clipped to a finger to measure blood oxygen levels, used in billions of clinical decisions every year. Multiple peer-reviewed studies, including findings acknowledged by the FDA, confirmed that pulse oximeters are significantly less accurate in patients with darker skin tones. The devices were calibrated on lighter-skinned subjects, and the engineering choices made during their development were never revisited at scale.
The consequences were not theoretical. Clinicians relying on pulse oximeter readings were, in some cases, receiving falsely reassuring readings for Black and Brown patients whose oxygen levels were dangerously lower than the device indicated. Patients who should have been escalated to intensive care were not. This is bias encoded in hardware, in a device so routine it barely registers as a "technology" anymore. And it persisted for decades before it was formally acknowledged as a patient safety issue.
"After watching a pulse oximeter struggle with my father's dark skin before he died from Covid, making health care AI more trustworthy became my mission."
— Amber Nigam, STAT News, November 2024
The pulse oximeter story is important not because it is uniquely egregious, but because it illustrates how bias can be invisible for decades when the populations most affected have the least power to surface it. Who gets to flag a product failure? Who gets to be heard when they do? Those are not engineering questions. They are structural ones.
The Digital Divide in Telemedicine
Beyond algorithmic and device-level bias lies a third dimension of inequity that HealthTech often sidesteps entirely: access. Telemedicine expanded dramatically during the COVID-19 pandemic, and with good reason — it removed geographic and logistical barriers to care in ways that genuinely helped millions of people. But it also assumed something that is not universally true: that patients have reliable internet, a functioning device, and sufficient digital literacy to navigate a telehealth platform.
That assumption does not hold. The communities with the greatest need for expanded healthcare access — rural populations, elderly patients, low-income households — are also among those least likely to have the infrastructure that telemedicine requires. Broadband access in the United States remains deeply unequal along racial and geographic lines. A patient in a remote area without reliable internet cannot benefit from a platform that requires a stable video connection. A 70-year-old without smartphone experience is not made "more accessible" by an app that requires three levels of navigation to book an appointment.
Telemedicine, deployed without intentional equity design, does not democratise care. It creates a faster, more convenient lane for the already-connected, while the structurally excluded remain where they were — or fall further behind as in-person resources are quietly consolidated to accommodate the digital-first model.
Three layers of inequity in HealthTech — and where they originate:
01 — Algorithmic bias: Models trained on historically unequal data encode and scale that inequality into clinical decisions
02 — Device bias: Medical hardware calibrated on non-representative populations produces inaccurate readings for the populations excluded from its development
03 — Access inequity: Digital infrastructure gaps mean the patients with the greatest need are those least able to use the tools designed to help them
What Actually Fixing This Looks Like
I want to resist the impulse toward vague optimism here, because that is one of the ways the tech industry avoids accountability. Saying "we need more diverse data" is not a plan. Saying "we are committed to equity" without governance structures to enforce that commitment is marketing. What does it actually look like to build HealthTech that does not perpetuate — or deepen — existing inequities?
It starts at the data layer. Clinical AI models must be trained on datasets that are demographically representative, and the demographic composition of those datasets must be disclosed publicly as a matter of course — not as a voluntary addition but as a basic standard of transparency. The Oxford Open Digital Health study (2025) found that nearly half of US clinical AI models do not report the ethnic composition of their training data. That is not a gap. That is a choice, and it needs to be regulated accordingly.
It requires diverse development teams — not as a diversity initiative, but as a product quality intervention. Research consistently shows that teams with greater gender, racial, and socioeconomic diversity surface failure modes that homogenous teams miss. This is not an abstract moral argument. It is a straightforward engineering argument: if the people building a tool do not represent the populations the tool will affect, the tool will have predictable blind spots.
It requires equity audits before deployment, not after harm has accumulated. The current model — in which an AI tool is used widely until a study reveals its bias — is ethically indefensible. A 2024 review in the Annals of Internal Medicine proposed guiding principles for integrating health equity into AI development lifecycles, arguing that equity assessment must be embedded at every stage: data collection, model design, validation, and ongoing monitoring post-deployment. That framework should be the floor, not an aspirational ceiling.
And it requires honest engagement with the access problem. Building a telehealth platform without accounting for broadband deserts is not an oversight — it is a policy choice. Designing digital health tools without low-bandwidth modes, multilingual interfaces, and offline functionality is a choice to serve some patients and not others. Governments, health systems, and technology companies all have a role in closing the infrastructure gap that makes HealthTech innovation inaccessible to the communities that need it most.
The Moral Stakes
I have spent a lot of time in this blog thinking about the architecture of healthcare systems — the cryptographic protocols, the consent mechanisms, the security layers. Those things matter. But they are not the whole picture. A patient-controlled health record built on an impeccable blockchain is still a tool that serves only the patients the system chose to design for. Security and equity are not competing values. They are both preconditions for a healthcare system that is worth defending.
The World Health Organisation estimates that half the world's population lacks access to essential health services. HealthTech will not solve that alone — it would be naive to suggest otherwise. But it can either widen or narrow the gap. Right now, in too many domains, it is widening it: building faster and more sophisticated tools calibrated on narrow populations, and deploying them as if precision is the same as fairness.
It is not. Precision that excludes people is not progress. It is the same inequality, dressed in a more expensive interface.
What I am arguing for is not slower innovation. It is innovation that asks the harder questions from the start — questions about whose data trained this model, whose physiology calibrated this device, whose internet connection this service assumes. Those questions are not obstacles to building good technology. They are the definition of it.