A novel method for clinical risk prediction with low-quality data.

Area of Science:

• Computational biology
• Machine learning in healthcare
• Data science for clinical prediction

Background:

• Real-world clinical data often suffers from low quality, including missing values, feature redundancy, and class imbalance, hindering predictive model performance.
• Existing methods struggle to systematically address these data quality issues, often introducing noise through imputation or failing to resolve challenges.
• A robust framework is needed to improve the accuracy and reliability of predictive models for clinical risks.

Purpose of the Study:

• To propose a novel framework, AMI-Net3, designed to address low-quality data challenges in clinical risk prediction.
• To develop a method that can learn directly from incomplete data without relying on imputation, thus avoiding introduced noise.
• To effectively handle correlated/redundant features and imbalanced datasets in clinical risk modeling.

Main Methods:

• Patients are represented as bags of instances (feature-value pairs), mapped to an embedding space using a proposed feature embedding method.
• A multi-instance neural network incorporating multi-head attention and attention-based multi-instance pooling is employed to manage feature correlations and identify valuable information.
• Joint end-to-end optimization of feature embedding and the neural network, with auxiliary supervision and focal loss functions, is utilized to handle imbalanced labels.

Main Results:

• AMI-Net3 demonstrates effectiveness in clinical risk prediction tasks, including adverse drug reactions, schizophrenia relapse, and invasive fungal infections.
• The framework successfully mitigates the negative impact of missing data by learning directly from instances.
• Experimental results show the superiority of AMI-Net3 over competitive baseline methods on real-world datasets.

Conclusions:

• AMI-Net3 provides a systematic and effective approach to overcome data quality limitations in clinical risk prediction.
• The proposed method offers a robust solution for handling missing data, feature redundancy, and class imbalance in healthcare predictive modeling.
• This framework has the potential to significantly improve the reliability and accuracy of predictive models in real-world clinical settings.