Prediction Pipeline Selection for Incomplete Clinical Data via Missingness Fingerprints and Instance Augmentation.

Area of Science:

• Machine Learning
• Health Informatics
• Data Science

Background:

• Clinical prediction from electronic health records (EHRs) faces challenges due to missing data and limited labeled examples.
• Current methods require costly trial-and-error to select the best predictive pipeline for new datasets.
• Graph neural networks offer diverse strategies but lack a dominant architecture across different missingness patterns.

Purpose of the Study:

• To develop an automated algorithm selection method for clinical prediction pipelines.
• To overcome bottlenecks of instance scarcity and poor distance quality in meta-learning for EHR data.
• To enable efficient and accurate pipeline selection without extensive trial-and-error.

Main Methods:

• Constructive instance augmentation: Expanded 20 base EHR datasets to 83 meta-instances using controlled missingness injection and label trimming.
• Missingness fingerprint: Characterized each meta-instance with a 10-dimensional fingerprint.
• Dynamic-supervised metric learning: Optimized fingerprint feature weights using differential evolution for accurate pipeline recommendation.

Main Results:

• The metric-learned kNN recommender achieved the highest win rate (20.5%) among non-oracle strategies on the augmented dataset.
• The recommender successfully selected the correct pipeline more often than any fixed default.
• Cross-domain evaluation showed a regret of 0.025 (55% below random selection) when adapting the fingerprint module.

Conclusions:

• The proposed framework automates pipeline selection for clinical prediction from EHRs, addressing meta-learning limitations.
• The method efficiently recommends optimal pipelines using a static fingerprint, requiring no model sweeps at deployment.
• The framework demonstrates modularity and effectiveness across diverse clinical datasets and missingness patterns.