Predicting Hospital Readmissions: How Much EHR History Does AI Really Need?

Forum topic · 小凯 · 2026-05-04

Summary

A forum post discusses the paper "Temporal Data Requirement for Predicting Unplanned Hospital Readmissions" (arXiv: 2605.00738), which examines the optimal observation window for predicting 30-day readmissions after hip and knee replacement surgery. The study combines structured EHR data (over 4 million records: diagnoses, procedures, medications, lab results) with unstructured clinical notes (80,000+), processed using bag-of-words, TF-IDF, Doc2Vec, BERT, and ClinicalBERT. Key findings: recent data is more predictive than older history; unstructured clinical notes capture signals missing from structured data (e.g., poor patient compliance); there is an information saturation point beyond which adding more history yields no gains or even degrades performance due to noise; and the optimal time window varies by prediction task. The post concludes that medical AI needs not "all the data" but the right temporal window of relevant, recent, high-quality data, and offers four practical questions for builders of clinical prediction models.

Paper: Temporal Data Requirement for Predicting Unplanned Hospital Readmissions Authors: Ramin Mohammadi, Vahab vahdat, Sarthak Jain, Amir T. Namin, Ramya Palacholla, Sagar Kamarthi arXiv: 2605.00738 | 2026-04-30

The AI Doctor That "Reads Every Chart"

Imagine an AI system predicting whether a patient will be readmitted after joint replacement surgery. How much historical data should it look at — just the day of surgery, the past week, the past year, or three years of complete records? Intuition says "the more, the better." This paper shows that is not necessarily true.

The Temporal Window Dilemma in EHRs

As electronic health records (EHRs) proliferate, AI models can access massive historical data. Yet a key engineering question has been overlooked: what length of historical data yields the best prediction?

Window too short: insufficient information, poor accuracy
Window too long: more noise, higher computational cost, potentially outdated information
Where is the optimum?

The paper studies the optimal observation window for predicting 30-day readmissions after hip and knee replacement surgery.

Structured Data vs. Unstructured Text

Structured data (4M+ records): diagnosis codes, procedure records, medication records, lab results.

Unstructured text (80,000+ clinical notes): physician progress notes, nursing notes, discharge summaries, consult opinions.

Text processing tools: traditional methods (bag-of-words, TF-IDF) and deep learning methods (Doc2Vec, BERT, ClinicalBERT).

Core question: how does the predictive power of these features change across different temporal windows?

Findings: Not All History Is Equally Useful

1. Recent data matters more than distant data. Perioperative data is most predictive; chronic-disease records from three years ago contribute little to predicting 30-day readmission. 2. Unstructured text has unique value. Details in clinical notes (e.g., "poor patient compliance") are often absent from structured data; NLP models extract signals structured data cannot capture. 3. There is an "information saturation" point. Beyond a certain window, adding more history no longer improves performance — and may even degrade it due to noise. 4. The optimal window is task-dependent. Different prediction targets (30-day vs. 1-year readmission) require different windows; there is no one-size-fits-all answer.

Feynman-Style Judgment: Information Value Decays Over Time

> "Information is like fruit — most valuable when fresh. Stale information is not just useless; it can mislead."

In medical prediction:

> "A three-year-old hypertension record helps predict tomorrow's readmission far less than last week's infection markers. It's not that more data is better — it's that *relevant*, *recent*, *high-quality* data is better."

The deeper insight: in medical AI we need not just "big data," but "the right temporal window."

Takeaways for Practitioners

If you are building a clinical prediction model, ask yourself:

1. Am I blindly using all available historical data? 2. How much do data at different time scales contribute to my prediction target? 3. Am I fully exploiting unstructured text (e.g., clinical notes)? 4. Is there an "information saturation" point beyond which more data no longer helps?

The core lesson: in medical AI, "when" matters as much as "what." Not all history is worth remembering. A smart system knows which past signals truly predict the future and which are noise. For readmission prediction, AI needs not "a lifetime of charts" but "the right-sized window."