Healthcare AI: measuring what changes for the patient

I was catching up on healthcare news from the past week when I stumbled upon this: https://www.fiercehealthcare.com/finance/healthleap-banks-38m-funding-backed-sequoia-capital-first-round-capital-and-hummingbird - it brings attention to a practical application of healthcare AI: identifying patients whose clinical needs might otherwise be missed. It got me thinking about a broader question around healthcare AI implementations: once an AI model identifies a patient who needs attention, what does the organization do differently, and how will it know whether that change helped?

When we evaluate a model, we usually examine how well it predicts. During implementation, we may also track adoption, usage, and the number of alerts generated. These measures tell us whether the technology performs and whether people use it. Establishing patient benefit requires following the prediction further.

Following a prediction through care

Let’s consider a hypothetical hospital implementing a model to identify discharged patients at risk of readmission. All figures in this example are illustrative.

Each week, the model flags 100 patients. The care team successfully contacts 40. Of those, 25 have a follow-up appointment arranged, and 18 attend.

Which number should appear in the program’s results?

All four are useful, but they describe different things. The model generated 100 flags. The team reached 40 patients. Twenty-five appointments were arranged. Eighteen visits occurred.

Reporting only the first number would leave us unsure whether care changed. Reporting the 18 completed visits would provide evidence that part of the care pathway happened. We would still need to determine whether those visits addressed patients’ needs and improved relevant outcomes.

The patients who were not reached also matter. We need to understand whether they were already receiving appropriate follow-up, could not be contacted, declined, or remained unaddressed because of limited capacity.

How a technically successful implementation can disappoint

Suppose the program’s success target is to identify more high-risk patients. The model meets that target, and the organization expands its coverage.

But the care team’s capacity remains unchanged. More patients enter the worklist, while the number receiving follow-up barely moves.

The program has improved its ability to identify possible need without improving its ability to respond. If evaluation focuses on prediction and usage, this constraint can remain hidden behind encouraging results.

This is one route to implementation failure that concerns me. The model may perform as intended, while the service using it falls short of its clinical objective.

An outcome measure will not create additional capacity by itself. It does, however, keep the intended benefit visible and prompt the organization to investigate why it has not materialized.

Work backward from the intended patient benefit

We should define the benefit we expect before deciding which implementation measures to report.

For a readmission program, that might include fewer avoidable returns to hospital, alongside measures of safety, recovery, or patient experience. Clinical teams should help determine which outcomes the proposed care pathway can reasonably influence.

We can then work backward through the steps needed to produce that benefit. Who reviews the risk signal? What assessment or follow-up should occur? Who owns the next action? How quickly does relevant information need to arrive?

This also makes the data requirements more specific. An appointment booked is different from an appointment attended. A care plan documented is different from care delivered. A note signed in the evening may describe an assessment completed that morning.

Process measures help explain whether the pathway worked. Patient outcomes help establish whether it achieved its purpose.

Measuring outcomes requires more than a before-and-after chart

Suppose, in a separate hypothetical example, readmissions fall from 12% before implementation to 10% afterward.

Among 1,000 patients in each period, that would be 120 readmissions before and 100 afterward: 20 fewer observed readmissions.

It is an encouraging change. Attributing it to the AI program requires understanding what else changed.

The patients may have had different levels of illness. Staffing or discharge practices may have improved. Another care program may have started during the same period.

A suitable comparison group or phased rollout can strengthen the evaluation. Comparing patients who received follow-up with those who did not may be misleading, because clinical teams may deliberately prioritize the sickest patients.

We should also examine missing outcomes and whether benefits persist across patient groups and routine operating conditions. A program’s result should remain credible after the initial attention surrounding its launch has passed.

Keep financial value connected to the clinical claim

Financial results also need a clear interpretation.

Additional reimbursement can benefit a hospital. It does not automatically represent savings to a payer or improved health for a patient. Shorter stays may release capacity, but their value depends on whether patients leave safely, how that capacity is used, and whether expenses change.

For an investment decision, one should look at the costs of software, integration, monitoring, and additional care alongside the benefits being claimed.

The evaluation should match the purpose of the tool. An administrative application may primarily aim to reduce workload. A clinical AI program introduced to improve care needs evidence that addresses that clinical objective.

Decide what would justify expansion

A model-development study can legitimately focus on predictive performance, but an organization deciding to expand a clinical service needs to evaluate what happens when that model becomes part of everyday care. The practical question is: what changed for the patient because the organization used the AI, and whether we have measured that change well enough.