What cancer symptom prediction teaches us about longitudinal health

I recently read a JCO Clinical Cancer Informatics paper on predicting cancer symptom trajectories using longitudinal EHR data.

The models are interesting, but the lesson is structural. Prediction only works because symptom history is preserved across time. When EHRs reduce symptoms to isolated encounter entries, learning breaks down.

The paper uses sparse, irregular nursing documentation and still shows that symptom trajectories can be learned. That matters for anyone building health software in the real world, where data is never clean.

I wrote a longer reflection connecting this to longitudinal health records, health graphs, and EHR Lite systems here:
https://myaether.live/blog/predicting-cancer-symptom-trajectories-longitudinal-ehr

Would love thoughts from others building in healthcare or working with messy time-series data.