Predictive Analytics (Healthcare)
Also known as: clinical prediction models, risk stratification, healthcare forecasting, predictive modeling in healthcare, clinical predictive analytics
The use of statistical or machine learning models on clinical and operational data to estimate the likelihood of a future event, such as deterioration, sepsis onset, readmission, or a missed appointment.
Predictive analytics in healthcare spans clinical and operational uses that carry very different risk profiles. Clinical prediction, for example sepsis onset or inpatient deterioration, influences care decisions and therefore attracts the heaviest evidence and governance requirements, including scrutiny of what population the model was validated on. Operational prediction, such as no-show likelihood or staffing demand, generally sits outside clinical decision-making but can still produce equity problems when the prediction drives resource allocation. In both cases the question that most determines real-world performance is whether the validation population resembles the population the model will actually run on. The documented failures in this field bear that out: an externally validated proprietary sepsis model missed roughly two thirds of cases while alerting on about 18 percent of admitted patients, and a widely used population health algorithm encoded racial disparity by predicting healthcare cost as a proxy for healthcare need. Neither was a modeling error in the narrow sense. One was a validation population mismatch, the other was a choice of outcome variable.