Predictive modeling in urgent care: a comparative study of machine learning approaches

Area of Science:

• Clinical informatics
• Artificial intelligence in medicine
• Machine learning for healthcare

Background:

• Electronic health records (EHRs) provide rich data for medical research.
• Machine learning (ML) shows promise in extracting insights from clinical data.
• Urgent care settings present unique predictive modeling challenges.

Purpose of the Study:

• To systematically compare ML algorithms for predictive modeling in urgent care.
• To evaluate algorithm performance on tasks like mortality prediction and disease marker discovery.
• To leverage the Medical Information Mart for Intensive Care (MIMIC-III) database for robust analysis.

Main Methods:

• Assessed 4 benchmark prediction tasks using EHR data (medical history, time-series, demographics).
• Utilized the MIMIC-III database for comprehensive patient information.
• Evaluated performance using AUC, F-1 score, sensitivity, and specificity.

Main Results:

• Recurrent neural networks achieved high accuracy (AUC > 0.90) in mortality prediction.
• Temporal models offered no significant advantage over deep models for differential diagnostics.
• Readmission risk appears independent of patient discharge stability.
• A multiclass scheme improved length-of-stay prediction with outliers.

Conclusions:

• Recurrent neural networks are highly effective for in-hospital mortality prediction.
• ML models can be tailored for specific urgent care predictive tasks.
• Further research can refine ML applications for length of stay and readmission prediction.