A comparative study: classification vs. user-based collaborative filtering for clinical prediction.

Area of Science:

• Health Informatics
• Machine Learning
• Biomedical Data Science

Background:

• Recommender systems, particularly user-based collaborative filtering (CF), are valuable for personalized predictions.
• CF has been recently applied to clinical risk prediction, treating patients as individuals and clinical data as items.
• This approach leverages patient similarity and prior data for risk assessment.

Purpose of the Study:

• To assess the performance of recommender systems, specifically user-based CF, for clinical risk prediction.
• To compare CF against traditional machine learning models (logistic regression, random forests) using various data imputation strategies and missing data scenarios.
• To evaluate performance on diverse, real-world medical datasets and simulated data.

Main Methods:

• Recasting supervised learning problems as unsupervised learning for risk prediction.
• Comparing user-based CF with logistic regression and random forests.
• Utilizing four public medical datasets (NHANES, SUPPORT, CKD, dermatology) and simulated data.
• Analyzing performance under varying degrees of missingness (MAR, MCAR) and class imbalance.

Main Results:

• User-based collaborative filtering (CF) consistently underperformed logistic regression and random forests across all tested datasets and conditions.
• The performance of CF was inferior regardless of imputation methods or the amount of missing data.
• These findings highlight significant limitations of CF in clinical risk prediction tasks.

Conclusions:

• User-based collaborative filtering is generally not desirable for clinical risk prediction when traditional classification methods are viable.
• Caution is warranted when applying CF to clinical datasets, particularly in the context of "Big Data".
• Alternative applications for CF in "Big Data" scenarios are discussed, alongside reasons for its limitations in risk prediction.