Why implementing machine learning algorithms in the clinic is not a plug-and-play solution: a simulation study of a...

Area of Science:

• Medical Informatics
• Clinical Decision Support Systems
• Artificial Intelligence in Healthcare

Background:

• Artificial intelligence (AI) and machine learning (ML) show promise in clinical medicine but often lack real-world validation.
• Many published algorithms remain unvalidated in practical clinical settings, hindering widespread adoption.
• This study addresses the gap by simulating the implementation of a specific ML algorithm for acute leukaemia diagnosis.

Purpose of the Study:

• To simulate the practical implementation challenges of the AI-PAL ML algorithm for acute leukaemia diagnosis.
• To evaluate the real-world performance of AI-PAL in a clinical setting.
• To highlight the necessity of local validation and adjustments for ML algorithms in healthcare.

Main Methods:

• A detailed simulation of the AI-PAL algorithm's implementation was conducted at the University Hospital Essen.
• Cohort building utilized a Fast Healthcare Interoperability Resources (FHIR) database to identify relevant patient cases.
• Algorithm performance was assessed by reproducing prior study results and analyzing diagnostic outcomes.

Main Results:

• The AI-PAL algorithm exhibited significantly lower performance in the simulated clinical implementation compared to published results.
• Area under the ROC curve for acute lymphoblastic leukaemia was 0.67 and for acute myeloid leukaemia was 0.71.
• Recalibrating probability cutoffs increased confident acute leukaemia diagnoses from 98 to 160, indicating a need for local adjustments.

Conclusions:

• Implementing ML algorithms in clinical practice presents significant challenges.
• ML models like AI-PAL require substantial adjustments and recalibration for optimal performance in diverse clinical settings.
• Local performance validation of clinical decision support algorithms is essential before routine integration to ensure reliability and patient safety.