Hypothesis Generation via Interpretable Machine Learning: A Case Study on Risk Factors for Postradiation Therapy Lung
Zheng Zhang1, Sang Ho Lee2, Rich Caruana3,4
1Department of Radiation Oncology, University of Michigan, Ann Arbor, Michigan.
Advances in Radiation Oncology
|June 8, 2026
Summary
Explainable Boosting Machine (EBM) shows promise for generating hypotheses in early-stage lung cancer research, despite modest performance on limited data. Its interpretable nature aids in identifying prognostic factors and potential biases.
Area of Science:
- Oncology
- Machine Learning
- Biostatistics
Background:
- Interpretability is crucial for trustworthy oncologic outcome prediction, especially with limited data.
- Existing methods often rely on post-hoc explanations for black-box models.
- Explainable Boosting Machine (EBM) offers an intrinsically interpretable 'glass-box' alternative.
Purpose of the Study:
- Investigate the utility of EBM for hypothesis generation in early-stage lung cancer.
- Assess EBM's ability to identify prognostic features and potential biases.
- Compare EBM's performance and interpretability against traditional models.
Main Methods:
- Applied EBM to a lung cancer recurrence dataset post-radiation therapy.
- Compared EBM-identified features with univariate analysis results.
- Benchmarked EBM against logistic regression and random forest models.
Main Results:
- EBM identified primary tumor size and BMI as key prognostic features, consistent with univariate analysis.
- The model revealed potential age bias and race-BMI interaction confounding.
- EBM provided competitive performance and enhanced interpretability over logistic regression and random forest.
Conclusions:
- EBM's interpretability makes it valuable for hypothesis generation in limited-data oncology settings.
- Modest performance currently limits its use as a clinical decision support tool.
- Further research is needed to address generalizability challenges with small datasets.
Related Concept Videos
01:25Statistical Methods for Analyzing Epidemiological Data
Epidemiological data primarily involves information on specific populations' occurrence, distribution, and determinants of health and diseases. This data is crucial for understanding disease patterns and impacts, aiding public health decision-making and disease prevention strategies. The analysis of epidemiological data employs various statistical methods to interpret health-related data effectively. Here are some commonly used methods:
01:21Cancer Survival Analysis
Cancer survival analysis focuses on quantifying and interpreting the time from a key starting point, such as diagnosis or the initiation of treatment, to a specific endpoint, such as remission or death. This analysis provides critical insights into treatment effectiveness and factors that influence patient outcomes, helping to shape clinical decisions and guide prognostic evaluations. A cornerstone of oncology research, survival analysis tackles the challenges of skewed, non-normally...
01:23Types of Biopharmaceutical Studies: Controlled and Non-Controlled Approaches
Biopharmaceutical studies constitute a vital field aiming to enhance drug delivery methods and refine therapeutic approaches, drawing upon diverse interdisciplinary knowledge. In research methodologies, the choice between controlled and non-controlled studies significantly influences the study's reliability and accuracy.
Non-controlled studies, commonly employed for initial exploration, lack a control group, rendering them susceptible to biases and external influences. In contrast, controlled...
Non-controlled studies, commonly employed for initial exploration, lack a control group, rendering them susceptible to biases and external influences. In contrast, controlled...