Natural Disease Progression

After a hospital discharge, only a limited number of people may receive intensive follow-up. Can their existing medical histories help identify those most likely to have a serious condition recorded during the next year?

This task predicts risks for conditions such as heart failure, stroke and kidney disease at thirteen horizons from 30 days to one year. Its central question is practical: if follow-up reaches the highest-risk 5% of eligible patients, what fraction of future cases does that group contain?

The model reads diagnoses, laboratory results, prescriptions, medication administrations and demographics available by discharge. A specimen collected earlier but reported later cannot be used: doing so would give the model information unavailable at the real decision time.

About 68,000 patients support training, with separate validation and hidden test patients. Earlier records use an older disease-coding system, creating imperfect labels when codes cannot be mapped cleanly. A missing later diagnosis can also mean that care happened elsewhere. The target is a recorded inpatient diagnosis, not the biological moment a disease began.

The baseline combines two views of a history. One summarizes which events occurred; a recurrent neural network reads events in order and updates a compact memory. Age and sex provide additional context. Together these views can distinguish, for example, a persistently abnormal measurement from a recent change.

The model produces a risk trajectory for each endpoint. Risk of an event occurring by a later date cannot be lower than risk by an earlier date, so the prediction mechanism accumulates nonnegative increments over time.

Training gives extra weight to rare disease events and excludes outcomes whose observation is incomplete. This avoids teaching the model that every missing follow-up record means good health.

The baseline was selected using overall ranking quality, whereas the task rewards detection within a small follow-up group. A model can rank cases reasonably well overall yet place too few in its highest-risk 5%. Aligning those goals is a concrete research opportunity.

Imagine 100 eligible patients, including 20 who later have the condition recorded. If the five highest-ranked patients include four of those cases, recall is 4/20, or 20%. Random selection would find about 5% on average. The real score averages this calculation across prespecified disease endpoints.

Patients with an existing endpoint or direct precursor are excluded from that endpoint’s assessment. Ranking and probability accuracy are checked separately: finding high-risk patients does not establish that a prediction of “20% risk” is numerically trustworthy.

The temporal neural model finds about 24% of future cases within the 5% alert budget, compared with about 5% for a constant-risk control. A tree-based research reference reaches about 27%; it is a useful comparison but not an eligible final model for this neural-model task.

The neural model’s probabilities remain imperfectly calibrated. These are single-run results, and validation and test averages cover different eligible endpoints. They support further study of prioritization, not evidence that using the system would improve patient outcomes.

A useful comparison would remove event order while preserving event content. Another would remove laboratory and medication information. These experiments test whether the model benefits from changing physiology or mainly from how often a patient is documented. Evaluate case detection and calibrated risks together, especially across coding eras.