Combining Machine Learning With Real-World Data to Identify Gaps in Clinical Practice Guidelines: Feasibility Study...

Area of Science:

• Neurology
• Health Informatics
• Medical Guidelines

Background:

• Clinical practice guidelines (CPGs) are crucial for patient care but often not fully implemented.
• Acute ischemic stroke requires timely treatment, making guideline adherence critical for outcomes.
• Gaps between CPGs and clinical practice can lead to treatment variability, especially when guidelines don't cover all scenarios.

Purpose of the Study:

• To systematically identify and quantify discrepancies in German thrombolysis-in-stroke guidelines.
• To utilize real-world data and machine learning (ML) for uncovering guideline gaps.
• To refine CPGs by identifying factors influencing clinical decisions.

Main Methods:

• Analysis of 13,440 patients from the German Stroke Registry - Endovascular Treatment (GSR-ET) (2015-2023).
• Development of a random forest model to predict thrombolysis treatment decisions using guideline-recommended, clinician-selected, and registry data features.
• Interpretation of model feature importance (permutation importance, Shapley values) to identify guideline deviations.

Main Results:

• The ML model achieved strong predictive performance (AUC 0.71-0.77) across different feature sets.
• Time from symptom onset to admission was the primary predictor for thrombolysis decisions.
• Patient age, not considered in current German guidelines, significantly predicted treatment decisions in the real-world data.

Conclusions:

• An innovative ML approach using real-world data effectively identified gaps between CPGs and clinical practice in stroke thrombolysis.
• Factors like patient age and pre-stroke status may influence treatment decisions beyond current guideline recommendations.
• This methodology can inform future refinements of clinical practice guidelines for acute ischemic stroke.