How Much Medical History Does AI Need to Predict Hospital Readmissions?

Forum topic · 小凯 · 2026-05-04

Summary

A paper by Ramin Mohammadi, Vahab Vahdat, Sarthak Jain, Amir T. Namin, Ramya Palacholla, and Sagar Kamarthi (arXiv:2605.00738) investigates the optimal temporal observation window for predicting unplanned 30-day readmissions after hip and knee replacement surgery. Using over 4 million structured EHR records (diagnoses, procedures, medications, labs) and more than 80,000 unstructured clinical notes processed with BOW, TF-IDF, Doc2Vec, BERT, and ClinicalBERT, the study shows that more history is not always better. Recent perioperative data carries the most predictive signal, clinical notes contain information absent from structured codes, and performance saturates—or degrades—beyond a certain window length due to noise and outdated information. The optimal window also varies by prediction target, meaning there is no one-size-fits-all answer. The takeaway for healthcare AI practitioners: data recency and relevance matter more than volume, and choosing the right temporal window is as important as choosing the right features.

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

Overview

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 records? Intuition says "the more, the better," but this paper shows that is not necessarily true.

The Temporal Window Dilemma in EHRs

With the spread of electronic health records (EHRs), AI models can access massive historical data. Yet a practical engineering question is often overlooked: how long a history yields the best predictions?

Too short a window: insufficient information, poor accuracy
Too long a window: added noise, higher compute cost, potentially outdated information
Where is the optimum?

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

Structured Data vs. Unstructured Text

Structured data (4M+ records): diagnosis codes, procedure records, medications, lab results
Unstructured text (80k+ clinical notes): physician progress notes, nursing notes, discharge summaries, consults
NLP tools: bag-of-words (BOW), TF-IDF, Doc2Vec, BERT, ClinicalBERT

The core question: how does the predictive power of these features change across different time windows?

Key Findings

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

Takeaway

In medical prediction, information value decays with time: a hypertension record from three years ago helps far less in predicting tomorrow's readmission than last week's infection markers. It is not about having more data, but data that is *relevant*, *recent*, and *high quality*.

If you are building clinical prediction models, ask yourself:

1. Am I blindly using all available history? 2. How much do different time scales contribute to my target? 3. Am I fully exploiting unstructured clinical notes? 4. Is there a saturation point beyond which more data no longer helps?

The core message: in healthcare AI, *when* matters as much as *what*. Not all history is worth remembering — a smart system knows which past signals predict the future and which are just noise. AI needs not "a lifetime of records," but "the right window."

> Source: https://arxiv.org/abs/2605.00738