A Practical Approach to Disease Risk Prediction: Focus on High-Risk Patients via Highest-k Loss

Area of Science:

• Healthcare AI and Machine Learning
• Predictive Analytics in Medicine
• Computational Health Sciences

Background:

• Disease risk prediction models are crucial for proactive healthcare but struggle with practical application due to resource limitations for high-risk patient follow-up.
• Existing models often fail to adequately address the challenge of minimizing false positives among high-risk individuals, impacting resource allocation and intervention effectiveness.

Purpose of the Study:

• To propose a novel and practical approach, Highest-k Loss, to enhance disease risk prediction by minimizing false positives specifically within high-risk patient groups.
• To develop a method that prioritizes interventions for patients with the highest predictive scores, optimizing the use of limited medical resources.

Main Methods:

• Introduced the Highest-k Loss function, which estimates weights for the highest predicted scores using a differentiable sorting operation.
• Applied the Highest-k Loss to a diabetes prediction task using a large dataset (253,680 responses) from a U.S. health survey.
• Utilized nested cross-validation and an aggregated model on an independent test set for rigorous evaluation.

Main Results:

• The Highest-k Loss significantly improved precision (positive predictive value) for the top 1%, 5%, and 10% of predicted scores compared to traditional Binary Cross Entropy and Focal Loss.
• Specific precision improvements included 0.05 for the highest 1% (95% CI: 0.041-0.055), 0.03 for the highest 5% (95% CI: 0.024-0.032), and 0.02 for the highest 10% (95% CI: 0.016-0.021).
• Demonstrated a practical solution for risk prediction that focuses on actionable patient cohorts rather than the entire population.

Conclusions:

• The Highest-k Loss function offers a practical and effective solution to a key limitation in current disease risk prediction models.
• This method enhances the ability to identify and focus on genuinely high-risk patients, thereby optimizing healthcare resource allocation and intervention strategies.
• The approach provides a valuable tool for improving the real-world applicability of predictive models in clinical settings.