Clinician-driven automated data preprocessing in nuclear medicine AI environments.

Area of Science:

• Clinical informatics
• Artificial intelligence in medicine
• Machine learning for healthcare

Background:

• Artificial intelligence (AI) in clinical science necessitates extensive data preprocessing (DP), traditionally driven by data scientists using mathematical rules.
• Clinician involvement is crucial for effective DP, yet current methods often lack direct clinical input, leading to suboptimal AI model development.
• This study introduces a novel DP approach prioritizing clinical domain knowledge to make AI model development more inclusive for clinicians.

Purpose of the Study:

• To propose and evaluate a clinician-driven data preprocessing approach for artificial intelligence models in clinical science.
• To demonstrate the impact of integrating explicit and non-explicit clinical rules into the data preprocessing pipeline.
• To enhance the inclusivity of the data preprocessing phase for clinical domain experts.

Main Methods:

• Introduced the Rule Set Table (RST) as an interface for clinicians to input formal rules (exp-keep, exp-remove, pref-keep, pref-remove features/samples) in a human-readable format.
• Integrated RST with common preprocessing algorithms for diverse clinical cohorts (prostate, glioma, DLBCL) in single and multi-center settings.
• Evaluated the impact of RST using Monte Carlo cross-validation and XGBoost for classification, comparing performance with and without RST in manual and automated DP setups.

Main Results:

• Machine learning (ML) models using manual preprocessing combined with RST achieved up to an 18% increase in balanced accuracy (BACC) compared to models without RST.
• The 'exp-keep' and 'pref-keep' instructions within RST demonstrated the highest performance gains: +18% BACC (glioma), +6% BACC (prostate), and +3% BACC (DLBCL).
• The study confirmed the added value of RST in improving the predictive performance of oncology-specific ML models across different datasets and scenarios.

Conclusions:

• The Rule Set Table (RST) effectively integrates clinical domain knowledge into the data preprocessing pipeline, enhancing AI model performance.
• This clinician-driven approach serves as a proof of concept for more inclusive data preprocessing in future clinical AI studies.
• The findings highlight the significant benefit of incorporating clinical expertise directly into the development of AI tools for healthcare.