Predictive analytics in healthcare means applying statistical or machine learning models to clinical and operational data to estimate the likelihood of a future event. That definition covers two activities with almost nothing in common from a risk perspective, and treating them as one category is the first error most organizations make.
Clinical prediction estimates something about a patient's body: sepsis onset, inpatient deterioration, readmission risk, mortality risk. It influences care decisions, so it carries the heaviest evidence and governance burden. Operational prediction estimates something about the delivery system: appointment no-shows, staffing demand, bed capacity, supply consumption. It generally sits outside clinical decision-making, but it can still produce equity problems the moment the prediction starts allocating resources.
What Healthcare Predictive Models Actually Predict
| Prediction target | Who acts on it | Decision it drives | Risk tier |
|---|---|---|---|
| Sepsis onset | Bedside nursing, rapid response, hospitalists | Escalation, cultures, antibiotics, level of care | Highest. A missed case and an alert flood are both harmful. |
| Inpatient deterioration | Nursing, rapid response teams | Increased monitoring, transfer to higher acuity | High. Directly changes clinical attention. |
| 30 day readmission risk | Care management, discharge planning, population health | Transitional care, follow-up intensity, resource allocation | High. Resource allocation decisions carry equity exposure. |
| Disease onset or progression risk | Primary care, population health, payers | Screening intensity, outreach, care management enrollment | High. Drives who receives extra attention. |
| Appointment no-show likelihood | Scheduling, clinic operations | Overbooking, reminder intensity, slot management | Moderate. Low clinical risk, real equity risk if it gates access. |
| Staffing and census demand | Nursing administration, operations | Shift planning, float pool deployment | Lower. Operational error rather than clinical error. |
| Denial or payment risk | Revenue cycle | Claim review, documentation prompts, appeal prioritization | Moderate. Financial, with documentation integrity exposure. |
Notice that risk tracks the decision, not the technique. The same modeling method applied to no-show prediction and sepsis prediction produces wildly different governance obligations, because one changes a scheduling slot and the other changes whether a clinician walks into a room.
The Two Failures Worth Knowing in Detail
Healthcare predictive analytics has an unusually well documented failure literature. Two cases are worth knowing precisely, because between them they cover almost every way these systems go wrong.
The Epic Sepsis Model: When Reported Performance Does Not Travel
A proprietary sepsis prediction model deployed at hundreds of United States hospitals was externally validated by University of Michigan researchers and published in JAMA Internal Medicine in 2021. The external result was substantially worse than the performance reported by the developer. The model failed to flag roughly two thirds of patients who developed sepsis, while still generating alerts on about 18 percent of all hospitalized patients.
Both halves of that result matter. Missing two thirds of cases means the tool cannot be relied on as a safety net. Alerting on nearly one in five admitted patients means it imposes a real attention cost on nursing staff, which is how alert fatigue develops and how genuine alerts start getting dismissed. The lesson is not that the vendor was dishonest. It is that performance reported on a development population is a claim about that population, and it does not automatically travel to a different case mix, prevalence, or documentation culture.
The Cost Proxy: When the Model Predicts the Wrong Thing Correctly
Obermeyer and colleagues, publishing in Science in 2019, examined a widely used population health algorithm and found that Black patients assigned the same risk score as White patients were measurably sicker. The algorithm was not using race as an input. It was using healthcare spending as a proxy for healthcare need, and because less money is historically spent on Black patients at equivalent levels of illness, the proxy encoded that disparity directly into the risk score. The researchers estimated the bias reduced the number of Black patients identified for extra care by more than half.
This failure is harder to catch than the sepsis case, because the model was working. It predicted future cost accurately. The defect was in the choice of outcome variable, a decision made early in development and rarely revisited during procurement. Any model whose target is a proxy for the thing you actually care about deserves this scrutiny: ask what the model literally predicts, then ask whether that is the same as what you want to know.
Models Degrade Without Anyone Touching Them
Validation is a snapshot, not a guarantee. Model drift covers two related effects. Data drift occurs when the distribution of inputs shifts because the patient population, documentation habits, or measurement practices changed. Concept drift occurs when the relationship between inputs and outcome changes, for example after a new clinical protocol alters how a condition is treated.
Vendor-side model updates create the same problem from the opposite direction: performance changes because the product changed, on the company's release schedule rather than yours. Treat a vendor model update as a trigger for re-evaluation, and require notice of material changes in the contract. A recurring post-deployment performance review is the single governance commitment most likely to catch a real-world failure before a patient does.
Most Clinical Predictive Models Are Not FDA Regulated
This surprises buyers regularly. A large share of predictive decision support running inside electronic health records falls outside FDA device oversight, because clinical decision support meeting certain criteria is excluded from the device definition. A model can influence care decisions at hundreds of hospitals without ever having been reviewed as a medical device. Whether a specific product is regulated depends on its intended use, not on how sophisticated it is, which is covered further in software as a medical device.
The practical consequence is that the absence of FDA authorization tells you very little, and its presence tells you only about the specific authorized use. Neither substitutes for local validation.
The Disclosure Leverage Most Buyers Do Not Use
There is a federal transparency requirement that applies here and is widely underused. Under the ONC HTI-1 final rule, certified health IT that enables or interfaces with predictive decision support interventions must make 31 source attributes available, compared with 13 for evidence-based interventions. Developers were required to meet the criterion by December 31, 2024, with ongoing maintenance obligations from January 1, 2025 and annual attestation that the documentation has been reviewed and updated.
Those attributes cover the details and output of the intervention, its purpose, cautioned out-of-scope uses, development details and input features, the process used to ensure fairness in development, the external validation process, quantitative performance measures, ongoing maintenance, and the schedule for updates and continued validation or fairness assessment. ONC also points organizations to a FAVES framing, meaning fair, appropriate, valid, effective, and safe, as an evaluation approach.
What to Establish Before Deployment
- What the model literally predicts, stated as an outcome variable, and whether that is the same as the thing the organization actually cares about.
- The validation population: size, prevalence, demographics, care setting, and time period, compared against the population the model will run on.
- Whether validation was prospective in live workflow or retrospective on historical records, and what the difference was.
- Performance at the operating threshold you will actually use, including positive predictive value and expected alert volume per unit per shift, not only an area under the curve.
- Subgroup performance, reported separately rather than averaged, for the populations the organization serves.
- The named clinical owner, the monitoring cadence, the thresholds that trigger review, and the written conditions under which the model gets switched off.
- Notice obligations for vendor model updates, and what re-validation follows one.
The Operational Side Deserves Its Own Guardrail
Operational prediction gets governed loosely because it is not clinical, and that is usually right on safety and usually wrong on equity. A no-show model that drives reminder intensity is harmless. The same model gating access to appointment slots, or triggering overbooking that produces longer waits for the patients most likely to be predicted as no-shows, distributes a burden along lines that correlate with transportation, work flexibility, and income.
The guardrail is simple to state: if an operational prediction changes what a patient receives rather than what the organization prepares for, govern it as though it were clinical.
Related Reading
For vendor and model selection, see healthcare predictive analytics tools and models. For the company layer, see the clinical decision support and predictive risk category. For the governance structure around these commitments, see the AI governance framework for hospitals and the state healthcare AI laws tracker, since several states now require that adverse determinations rest on individual clinical evidence rather than group data.