Prediction of 30-Day All-Cause Hospital Readmissions Using Limited Structured Electronic Health Record Data:...

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Clinical Operations

Background:

• Unplanned hospital readmissions pose significant financial and operational challenges in the US healthcare system.
• Existing readmission prediction models often require extensive data available late in patient encounters, limiting real-time intervention.
• This creates a trade-off between model accuracy and timeliness for practical application.

Purpose of the Study:

• To determine if a limited set of structured clinical codes contains a clinically meaningful predictive signal for 30-day all-cause readmissions.
• To evaluate if predictive accuracy is retained when using restricted clinical code feature sets.

Main Methods:

• A retrospective study utilized a deidentified electronic health record dataset of 50,000 inpatient encounters.
• Two feature sets were created: a limited set (first 5 ICD-10 and CPT codes, CCI) and a rich set (all available codes plus CCI).
• Four models (random forest, CatBoost, multilayer perceptron, DistilBERT) were trained and evaluated using AUROC and F1-score on a hold-out set.

Main Results:

• Models using the limited feature set achieved comparable performance to those using the rich feature set.
• The best limited model (random forest) had an AUROC of 0.5596, closely matching the best rich model (DistilBERT) at 0.5703.
• DistilBERT on the limited feature set yielded the highest F1-score of 0.3434.

Conclusions:

• Models employing a limited set of structured clinical codes can achieve predictive performance comparable to models using more comprehensive coding information.
• This suggests that efficient, timely readmission prediction is feasible using readily available clinical data.
• The findings support the development of practical, real-time readmission risk assessment tools.