TL;DR
This paper introduces a novel machine learning pipeline, GBDT-SHAP, that efficiently identifies potential risk factors for mortality from massive biomedical datasets. By combining Gradient Boosting Decision Trees (GBDT) with SHAP (SHapley Additive exPlanations) for feature attribution, the method effectively handles non-linear relationships and interactions, offering a pragmatic first step for hypothesis-free risk factor discovery.
What Problem Does This Paper Address?
Large-scale biomedical datasets like the UK Biobank are treasure troves of information, holding immense potential for discovering new risk factors for diseases and mortality. However, traditional statistical methods, such as logistic and Cox regression, struggle with the sheer volume and complexity of these datasets. They are limited in the number of variables they can handle, assume linear relationships, and require careful modeling to account for interactions. This often leads researchers to pre-select a limited set of "known" or biologically plausible predictors, potentially missing novel associations hidden within the vast majority of available data. The challenge, therefore, is to develop an efficient and scalable approach for hypothesis-free risk factor discovery that can navigate this high-dimensional data landscape, uncovering potential associations that might otherwise remain undetected.
Methods
The core of this paper is the proposed GBDT-SHAP pipeline. Here's a breakdown of the key components:
- Gradient Boosting Decision Trees (GBDT): This ensemble learning method was used to build predictive models for mortality. GBDT is chosen for its ability to handle large datasets, manage missing values, and capture complex non-linear relationships and interactions between variables without extensive pre-processing. The authors specifically used the CatBoost implementation.
- SHAP (SHapley Additive exPlanations): To understand which predictors are most important for the GBDT model's predictions, SHAP values were employed. SHAP provides a unified measure of feature importance by attributing the contribution of each feature to the model's output. This allows for a more nuanced understanding of variable influence compared to traditional feature importance metrics.
- Feature Selection:
- SHAP Value Thresholding: Predictors with SHAP values above a certain threshold (≥ 0.05) were initially selected as potentially important.
- Correlation-Based Elimination: Highly correlated predictors were identified using Spearman's rank correlation (ρ > 0.9), and only one from each highly correlated pair was retained to avoid multicollinearity.
- Confounder Adjustment and Validation: After identifying potential risk factors using the GBDT-SHAP pipeline, Cox regression models were used for confounder adjustment and further validation. This step allowed for interpretability and assessment of associations while controlling for known confounders like age, sex, and socioeconomic status. False Discovery Rate (FDR) correction was applied to account for multiple testing.
The pipeline also incorporated learning curve analyses to ensure sufficient data for training and to assess model performance stability.
Datasets
The study utilized the UK Biobank dataset, a large-scale biomedical research resource.
- Participants: Over 502,506 individuals aged 37-73 years at recruitment.
- Follow-up: Approximately seven years for mortality registrations.
- Predictors: Initially, over 11,639 predictors were considered, encompassing a wide range of data including baseline characteristics, sociodemographics, lifestyle factors, physical measurements, cognitive function, psychosocial factors, self-reported diseases, medications, operations, health and medical history, and hospital diagnoses.
Key Insights and Ideas
- Hypothesis-Free Discovery in High Dimensions: The GBDT-SHAP pipeline demonstrates a powerful and pragmatic approach to sift through thousands of variables to identify potential risk factors without prior hypotheses. This is crucial for exploring novel associations in complex datasets.
- Capturing Non-Linearity and Interactions: GBDT's inherent ability to model non-linear relationships and interactions, combined with SHAP's interpretability, allows for the discovery of associations that traditional linear models might miss.
- Complementary Strengths: The study successfully integrates the predictive power of machine learning (GBDT) with the interpretability and validation capabilities of conventional statistical methods (Cox regression). This hybrid approach leverages the best of both worlds.
- "Hidden" Predictors Revealed: The pipeline was effective in identifying relevant predictors that might have been overlooked by traditional methods or by relying solely on pre-selected variables.
- Importance of Data Linkages: The study highlights the value of comprehensive data, including linkages to hospital diagnoses and cancer registries, which contributed a significant portion (49.6%) to the identified variable importance.
Key Results
- 193 Potential Risk Factors Identified: From an initial set of 11,639 predictors, 193 were identified as potential risk factors based on SHAP values and correlation tests.
- 166 Confirmed Associations: After adjusting for confounders using Cox regression, 166 out of the 193 identified predictors showed a statistically significant association with mortality (P < 0.05 after FDR correction).
- Predictor Categories: Hospital diagnoses accounted for the largest share of variable importance (49.6%), followed by health and medical history (10.2%), baseline characteristics (8.8%), and sociodemographics (8.8%). Lifestyle and environment factors also contributed significantly (6.5%).
- Confounding Identified: For 19 predictors, an association was observed in unadjusted analyses but disappeared after adjusting for confounders, indicating potential bias by confounding.
- Comparison with Other Methods: The GBDT-SHAP pipeline showed advantages over LASSO and XGBoost in terms of feature stability and the ability to capture important predictors, especially those with missing data or non-linear associations.
Limitations and Future Work
- Healthy Volunteer Bias: The UK Biobank cohort consists of volunteers with higher education and socioeconomic status, and lower mortality rates than the general population. This "healthy volunteer bias" may affect the external validity of the findings.
- Overidentification: As the study used a single dataset, the possibility of overidentification of predictors cannot be entirely ruled out.
- Missing Important Determinants: The exclusion of predictors with missing data for more than 5% of participants might have led to the omission of important determinants.
- Causality: The observational nature of the study and the limited adjustment for covariates mean that causal effects cannot be established, and confounding likely explains some associations.
- Interpretability of Interactions: While the pipeline can identify interactions, a more detailed modeling approach with domain expertise is still required for full interpretability of predictor-outcome associations.
The authors suggest that for future work, a more detailed modeling approach incorporating domain expertise and traditional methods is necessary to fully understand the identified predictor-outcome associations. They also emphasize the need for replication and further investigation of apparent associations, especially in the context of confounding and multicollinearity.
Summary updated