岡 宗一

Liu et al. (2025) analyzed UK Biobank data, using Principal Component Analysis (PCA) to identify lipid patterns associated with depression and bipolar disorder. Their work reported that the first principal component (PC1), reflecting Apolipoprotein B (ApoB), cholesterol, and low-density lipoprotein cholesterol (LDL-C), showed a protective effect against depression. However, their methodological approach warrants discussion. PCA is a linear dimensionality reduction technique. The authors noted nonlinear relationships between lipid profiles and mood disorder risk, contradicting PCA's inherent linearity assumption. Applying linear methods like PCA to nonlinear data can lead to significant distortions, systematic bias, and underfitting, failing to capture true data complexity. PC1 may have obscured genuine associations by forcing distinct biological features into a single linear equation, potentially diluting crucial signals. For future research, complementing PCA with unsupervised learning techniques like Feature Agglomeration (FA) and Highly Variable Gene Selection (HVGS) could offer a more robust approach. Additionally, using nonlinear nonparametric statistical methods such as Spearman's rho or Kendall's tau would be beneficial. These methods detect monotonic relationships without linearity assumptions, precisely capturing potentially nonlinear associations and enhancing interpretability in translational biomarker research.

Li et al. (2025) highlighted Random Forest's (RF) high accuracy and SHapley Additive exPlanations (SHAP)-derived feature importance for almond deterioration. However, concerns persist regarding the reliability of these interpretations, as high predictive accuracy doesn't guarantee valid feature rankings due to inherent biases in tree-based models, further amplified by SHAP's model dependency. To mitigate this, integrating robust statistical methods such as Spearman's rho, Kendall's tau, Total correlation and Effective transfer entropy is crucial for unbiased assessment. This combined approach ensures a more reliable evaluation of key indicators. Future research should prioritize methodologies combining machine learning with rigorous statistical validation for more interpretable and trustworthy insights in complex biological systems. This integrated approach holds significant promise for improving the reliability of feature importance evaluations, leading to more trustworthy insights applicable to food science and chemistry fields.

In medical machine learning (ML), a fundamental methodological distinction exists between optimizing model performance for predictive tasks and pursuing causal inference for mechanistic interpretation. Achieving high predictive accuracy does not necessarily imply that a model can uncover the true physiological mechanisms underlying the data. This letter addresses a critical interpretational challenge in medical machine learning, building upon Yuyang Yan et al.'s valuable work on exacerbation classification in asthma and COPD. While their multi-feature fusion model, particularly comprising models such as K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forest (RF), and Bidirectional Long Short-Term Memory (BiLSTM) demonstrates high predictive accuracy for respiratory exacerbations, we highlight that such performance alone does not guarantee reliable insights into feature importance. Complex tree-based models like RF, when interpreted via methods like SHapley Additive exPlanations (SHAP), can exhibit inherent biases, overemphasizing features used in early splits and reflecting what is important for their specific prediction rather than the true underlying physiological drivers. Validating feature importance remains challenging without ground truth, as different models often yield varying rankings. We argue that solely relying on model-dependent interpretations risks misrepresenting the actual mechanisms of complex medical phenomena. Therefore, we advocate for a robust analytical strategy that transcends mere predictive metrics. This involves a synergistic approach combining the predictive power of ML with impartial, complementary statistical methodologies-such as non-parametric correlation and mutual information-to ensure genuinely trustworthy scientific insights into the true drivers of respiratory exacerbations.

Liu et al. (2025) present an innovative approach to PM10 source apportionment in urban environments by integrating Positive Matrix Factorization with machine learning (ML) models including XGBoost, Random Forest (RF), and Support Vector Machine (SVM). Their use of the Lung Performance Optimization (LPO) algorithm for XGBoost and 10-fold cross-validation improved model robustness, with the LPO-XGBoost variant achieving the highest predictive accuracy (r2 = 0.88). SHAP values were employed to interpret feature importance, but concerns arise regarding the reliability of these rankings due to model-specific biases. Tree-based models may overemphasize features selected early in the decision process, while SVM models can obscure original feature relationships through kernel transformations. Although Liu et al. interpret variability in feature importance across models as analytical depth, this may reflect methodological inconsistencies rather than strength. SHAP values, being model-dependent, can inherit and amplify biases, complicating interpretation. In environmental research, where data are often noisy and high-dimensional, such instability can undermine the reliability of insights. Future studies should consider incorporating unsupervised learning techniques and non-parametric statistical methods to improve interpretability and robustness. Specifically, methods such as Feature Agglomeration (FA), Highly Variable Gene Selection (HVGS), Spearman's rho, and Kendall's tau can better capture complex and nonlinear associations, particularly in the context of health risk assessments. By integrating these approaches, researchers can enhance the stability of feature selection, reduce the influence of model-specific biases, and improve the transparency of analytical outcomes. A more systematic and cautious approach to model evaluation will ultimately strengthen reproducibility and support more informed environmental decision-making.

This correspondence critically examines the methodology of Schindele et al. (2025) on thyroid cancer recurrence prediction. While their interpretable XGBoost model achieved a high predictive accuracy of 95.8% and a 0.947 AUROC, it is crucial to recognize that this predictive power does not justify the reliability of its derived feature importance rankings. As widely acknowledged in the literature, high predictive accuracy does not guarantee unbiased or reliable feature attribution. We underscore that gradient boosting decision tree (GBDT) models, including XGBoost, are prone to inherent biases in feature importance estimation, often due to overfitting. Furthermore, SHapley Additive exPlanations (SHAP), a widely adopted explainable AI (XAI) technique, can inherit and even amplify these biases, given its model-dependent nature. This raises concerns about the interpretive validity of the identified risk factors. To mitigate these methodological limitations, we advocate for integrative analytical frameworks that combine machine learning with robust statistical and non-parametric approaches, such as Highly Variable Feature Selection (HVFS) and Independent Component Analysis (ICA). These multi-faceted strategies are indispensable for obtaining robust and interpretable insights into feature importance, warranting their prioritization in future research efforts.

This paper comments on the valuable contribution by Carvalho and Gavaia regarding machine learning for osteoporosis risk prediction, particularly their use of a stacking ensemble model and feature importance analysis. While acknowledging the model's high predictive accuracy, we raise a crucial concern: high accuracy does not inherently validate the reliability of feature importance interpretation. We discuss how the interpretation of feature importance from complex, model-dependent methods like those used can be influenced by model structure and data characteristics, potentially overemphasizing certain variables or reflecting model-specific relevance rather than true underlying causal drivers of osteoporosis risk. Validating feature importance is inherently difficult due to the absence of ground truth for causal relationships. To address these limitations and move beyond purely model-dependent predictive importance, we propose integrating complementary statistical methodologies, such as Spearman's rho, Kendall's tau, Mutual Information, and Total Correlation. These impartial and resilient methods can offer more robust insights into variable relationships. By combining predictive ML modeling with these statistical approaches, we aim to advance the understanding of complex health outcomes like osteoporosis in biomedical and healthcare applications, providing a more dependable assessment of feature importance and model behavior.

Pan et al. demonstrated the superior predictive performance of their machine learning ML models for soil phthalate PAE concentrations, highlighting the critical role of feature importance as assessed by SHapley Additive exPlanations (SHAP). Notably, the Multilayer Perceptron (MLP) model achieved the highest performance (R² = 0.8637), followed by SVR and XGBoost. However, concerns persist regarding the reliability of feature importance derived from these models and their SHAP interpretations. Specifically, predictive accuracy does not guarantee the validity of feature rankings due to the inherent biases present in tree-based, neural network, and kernel-based methods, which are further exacerbated by SHAP's inherent dependency on model outputs. To mitigate these biases, integrating robust statistical methods is crucial. Techniques such as Spearman's rho, Kendall's tau, Goodman-Kruskal's gamma, Somers' delta, and Hoeffding's dependence, combined with p-value analysis, offer unbiased assessments. Integrating these statistical methods alongside ML models ensures a more reliable evaluation of feature importance in environmental risk modeling. Consequently, future research should prioritize methodologies that combine ML with rigorous statistical validation to enhance accuracy and reduce biases.

Song et al. (2024), "Prediction of PFAS bioaccumulation in different plant tissues with machine learning models based on molecular fingerprints," employed machine learning methods, such as XGBoost and SHapley Additive exPlanations (SHAP), to predict PFAS bioaccumulation, reporting high predictive accuracy. However, this commentary critically examines their interpretation of feature importance, since high predictive accuracy does not guarantee reliable feature importance. Both XGBoost and SHAP are known to exhibit biases, such as overemphasizing features used in early splits and inheriting biases from the underlying model. Furthermore, the high dimensionality and potential collinearity of molecular fingerprints complicate SHAP interpretation, increasing overfitting risk and compromising SHAP value stability. To provide a general example, we conducted an independent simulation using a publicly available dataset of US industrial facilities and environmental compliance, demonstrating significant discrepancies between feature importance rankings from XGBoost and robust statistical tests. This commentary advocates for robust statistical methods coupled with p-values, including Spearman's rho, Kendall's tau, Goodman-Kruskal's gamma, Somers' delta, and Hoeffding's dependence, for feature selection. These non-parametric methods, which are independent of specific model assumptions and rely on data ranks, are better suited to capture complex relationships in high-dimensional data, providing a more reliable foundation for future PFAS bioaccumulation research.