Interpretable Machine Learning for Official Statistics

Interpretable vs. noninterpretable machine learning models for data-driven hydro-climatological process modeling

ObjectivePostoperative complications remain a major concern in laryngeal cancer surgery, often requiring invasive interventions or intensive care. This study aimed to develop and validate an interpretable machine learning (ML) model to preoperatively predict Clavien-Dindo Grade ≥ III complications and support risk-informed perioperative decision-making.MethodsWe conducted a retrospective study using a temporally split cohort of laryngeal cancer patients. Postoperative complications were graded using the Clavien-Dindo (CD) classification. Eight ML algorithms were trained and evaluated using receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA). Model interpretability was assessed using SHapley Additive exPlanations (SHAP). A web-based calculator was deployed for clinical use.ResultThe random forest (RF) model achieved the best performance, with an area under the curve (AUC) of 0.935 in the training set and 0.842 in the test set. The model demonstrated robust sensitivity and specificity for both surgical and medical complications. Calibration curves indicated strong agreement between predicted and actual outcomes. SHAP analysis identified eight key predictors—such as vocal cord mobility, tumor subsite, and nutritional status—that contributed most to risk estimation. A user-friendly web calculator was developed and is accessible at: https://qilushiny.shinyapps.io/qilupredicate/.ConclusionWe developed a clinically interpretable ML model that accurately predicts major postoperative complications in patients undergoing laryngeal cancer surgery. This tool provides individualized risk assessments that can guide surgical planning, optimize perioperative strategies, and enhance shared decision-making. Prospective multicenter validation is needed to confirm its utility in routine practice.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12893-025-03213-z.

ObjectiveDistant organ metastasis remains the primary factor affecting long-term survival following radical surgery for colorectal cancer (CRC). This study aimed to develop and validate an interpretable machine learning (ML) model to predict the 5-year cumulative risk of distant metastasis after radical CRC surgery.MethodsA retrospective observational cohort study was conducted using clinical and follow-up data from 341 CRC patients who underwent radical surgery. The cohort was randomly divided into a training set n=239 and a validation set n=102 at a 7:3 ratio. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, identifying variables associated with the 5-year cumulative occurrence of metastasis. Prediction models were constructed using seven algorithms. Model performance was evaluated through multiple metrics: area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, calibration plots, and decision curve analysis. The SHapley Additive exPlanations (SHAP) method was applied to improve model interpretability.ResultsLASSO combined with tenfold cross-validation selected 11 key features for model development. Among the models tested, the SVM model demonstrated superior performance, achieving a Brier score of 0.144 and an AUC of 0.865 in the validation set. Calibration and clinical decision curves confirmed the SVM model’s strong calibration and clinical applicability. The SHAP dependence plots and force analysis provided explanations at both feature and individual patient levels for the model’s 5-year risk predictions.ConclusionThis study established a high-accuracy and interpretable ML model capable of effectively predicting the 5-year cumulative risk of distant organ metastasis after radical colorectal cancer surgery, while further external validation is necessary to confirm its clinical utility.

Pediatric adrenocortical tumors (pACTs) are rare and clinically heterogeneous. Existing risk stratification systems rely on fixed thresholds and linear assumptions, which may limit their prognostic accuracy-particularly for nonmetastatic, locally advanced cases. We aimed to develop an interpretable machine learning (ML) model for individualized survival prediction using only routine clinical features. We retrospectively analyzed 97 patients with pACT from the German Pediatric Oncology Hematology-Malignant Endocrine Tumors Registry (1997-2024). An Extreme Gradient Boosting Cox proportional hazards model was trained using 4 features-tumor volume, distant metastases, pathologic T stage, and resection status-identified via systematic feature evaluation across 11 737 model combinations. Performance was assessed using a stratified 80/20 train-test split, 500 bootstrap iterations, and Harrell's concordance index (C-index). SHapley Additive exPlanations (SHAP) were used for interpretability. The model achieved strong prognostic performance (test-set C-index: 0.925; bootstrap mean: 0.891, 95% confidence interval: 0.817-0.946). SHAP analysis confirmed the dominant influence of metastatic status, followed by tumor volume, T stage, and resection status. The model uncovered nonlinear and additive effects, including a SHAP- and bootstrap-guided tumor volume cut-off (190 mL, 95% confidence interval 127-910 mL) that only slightly differed from conventional thresholds. Stratification remained robust in subgroups, including nonmetastatic patients with advanced local disease. This interpretable ML model enables individualized survival prediction in pACT using only routine clinical data. It offers a clinically accessible and clinically meaningful complement to existing scoring systems, particularly in patients with ambiguous risk profiles who may benefit from more personalized management.

Ovarian cancer (OC) is a deadly disease that affects a large number of women worldwide. Machine Learning (ML) models can help in the early detection of this disease, however, the use of these models may be limited by their lack of interpretability and the difficulty to evaluate their performance. In this work, five types of datasets were used, employing clinical features, different types of coding genomic features, and combining both. The use of interpretable ML (IML) models (one linear and one nonlinear model) provided us with better interpretability of the five feature sets. Following this study, nine binary classification models were compared, and the Accuracy, Recall, and Area Under the Curve were analyzed. The results showed that ML models employed the combination of clinical features and genomes with the coding of the position of genes in patients significantly improves the prediction. We demonstrated that the inclusion of different preprocessed patient data and especially through the information provided by IML models, can help clinicians to understand the disease better and make informed treatment decisions.

Repeat transurethral resection (ReTUR) is essential for reducing residual and recurrent non-muscle-invasive bladder cancer (NMIBC). Pathological upstaging after ReTUR significantly influences prognosis. This study aimed to develop an interpretable machine learning model using CT radiomics to predict the risk of pathological upstaging following ReTUR in NMIBC patients. We retrospectively analyzed 104 NMIBC patients who underwent ReTUR at the Second Affiliated Hospital of Nanchang University from March 2019 to July 2022. Data were split 7:3 into training and internal validation sets. An external validation set included 40 patients from two other hospitals. Radiomic features were extracted from preoperative CT scans. Least Absolute Shrinkage and Selection Operator (LASSO) and multivariate logistic regression were used to identify predictors of pathological upstaging. Four machine learning models, including Extreme Gradient Boosting (XGBoost), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Linear Discriminant Analysis (LDA), were constructed and evaluated using AUC, accuracy, precision, F1 score, calibration curves, and decision curve analysis (DCA). The best model was interpreted via SHapley Additive exPlanations (SHAP) to identify key predictive features. umor grade (OR = 7.02, 95% CI: 1.17-42.21), tumor size (OR = 5.83, 95% CI: 1.21-28.15), and tumor number (OR = 6.83, 95% CI: 1.18-39.52) were independent risk factors. From 4,738 radiomic features, nine were selected. The XGBoost model outperformed others, with an AUC of 0.804 (95% CI: 0.756-0.862), accuracy of 77.4%, precision of 82.7%, and F1 score of 0.701 in internal validation. External validation confirmed its robustness. SHAP analysis highlighted Wavelet_LLH_firstorder_Maximum.1, Gradient_ngtdm_Complexity, and tumor grade as top predictors. The model showed good calibration and clinical utility on DCA. An interpretable CT radiomics-based machine learning model integrating clinical and imaging features was developed to accurately predict pathological upstaging risk after ReTUR in NMIBC patients. This tool may support clinical decision-making for individualized treatment after multicenter validation.

The study aimed to develop and externally validate multiparametric MRI (mpMRI) radiomics-based interpretable machine learning (ML) model for preoperative differentiating between benign and malignant prostate masses. Patients who underwent mpMRI with suspected malignant prostate masses were retrospectively recruited from two independent hospitals between May 2016 and May 2023. The prostate mass regions in T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) MRI images were segmented by ITK-SNAP. PyRadiomics was utilized to extract radiomic features. Inter- and intraobserver correlation analysis, t-test, Spearman correlation analysis, and the least absolute shrinkage and selection operator (LASSO) algorithm with a five-fold cross-validation were applied for feature selection. Five ML learning models were built using the chosen features. Model performance was evaluated with internal and external validation, using area under the curve (AUC), calibration curves, and decision curve analysis to select the optimal model. The interpretability of the most robust model was conducted via SHapley Additive exPlanation (SHAP). A total of 567 patients were enrolled, consisting of the training (n = 352), internal test (n = 152), and external test (n = 63) sets. In total, 2,632 radiomic features were extracted from regions of interest (ROIs) of T2WI and DWI images, which were reduced to 18 via LASSO. Five ML models were established, among which the random forest (RF) model presented the best predictive ability, with AUCs of 0.929 (95% confidential interval [CI]: 0.885-0.963) and 0.852 (95% CI: 0.758-0.934) in the internal and external test sets, respectively. The calibration and decision curve analyses confirmed the excellent clinical usefulness of the RF model. Besides, the contributing relations of the radiomic features were uncovered using SHAP. Radiomic features from mpMRI combined with machine learning facilitate accurate preoperative evaluation of the malignancy in prostate masses. SHAP can disclose the underlying prediction process of the ML model, which may promote its clinical applications.

Improving Clinical Decisions in IR: Interpretable Machine Learning Models for Predicting Ascites Improvement after Transjugular Intrahepatic Portosystemic Shunt Procedures

BACKGROUND Acute kidney injury (AKI) is a common adverse outcome following nephrectomy. The progression from AKI to acute kidney disease (AKD) and subsequently to chronic kidney disease (CKD) remains a concern; yet, the predictive mechanisms for these transitions are not fully understood. Interpretable machine learning (ML) models offer insights into how clinical features influence long-term renal function outcomes after nephrectomy, providing a more precise framework for identifying patients at risk and supporting improved clinical decision-making processes. OBJECTIVE This study aimed to (1) evaluate postnephrectomy rates of AKI, AKD, and CKD, analyzing long-term renal outcomes along different trajectories; (2) interpret AKD and CKD models using Shapley Additive Explanations values and Local Interpretable Model-Agnostic Explanations algorithm; and (3) develop a web-based tool for estimating AKD or CKD risk after nephrectomy. METHODS We conducted a retrospective cohort study involving patients who underwent nephrectomy between July 2012 and June 2019. Patient data were randomly split into training, validation, and test sets, maintaining a ratio of 76.5:8.5:15. Eight ML algorithms were used to construct predictive models for postoperative AKD and CKD. The performance of the best-performing models was assessed using various metrics. We used various Shapley Additive Explanations plots and Local Interpretable Model-Agnostic Explanations bar plots to interpret the model and generated directed acyclic graphs to explore the potential causal relationships between features. Additionally, we developed a web-based prediction tool using the top 10 features for AKD prediction and the top 5 features for CKD prediction. RESULTS The study cohort comprised 1559 patients. Incidence rates for AKI, AKD, and CKD were 21.7% (n=330), 15.3% (n=238), and 10.6% (n=165), respectively. Among the evaluated ML models, the Light Gradient-Boosting Machine (LightGBM) model demonstrated superior performance, with an area under the receiver operating characteristic curve of 0.97 for AKD prediction and 0.96 for CKD prediction. Performance metrics and plots highlighted the model’s competence in discrimination, calibration, and clinical applicability. Operative duration, hemoglobin, blood loss, urine protein, and hematocrit were identified as the top 5 features associated with predicted AKD. Baseline estimated glomerular filtration rate, pathology, trajectories of renal function, age, and total bilirubin were the top 5 features associated with predicted CKD. Additionally, we developed a web application using the LightGBM model to estimate AKD and CKD risks. CONCLUSIONS An interpretable ML model effectively elucidated its decision-making process in identifying patients at risk of AKD and CKD following nephrectomy by enumerating critical features. The web-based calculator, found on the LightGBM model, can assist in formulating more personalized and evidence-based clinical strategies.

Optimisation and interpretation of machine and deep learning models for improved water quality management in Lake Loktak

ObjectiveThis study aimed to develop and validate interpretable machine learning (ML) models based on multi-dimensional fusion data for predicting positive surgical margins (PSM) in robot-assisted radical prostatectomy (RARP).MethodsPatients who underwent RARP at our institution between January 2016 and July 2025 were enrolled. Demographic, clinical, biopsy pathology data, and MRI-derived anatomical features (measured using ITK-SNAP on axial, sagittal, and coronal planes) were collected. Feature selection was performed using intraobserver and interobserver correlation coefficients (ICCs), low-variance filtering, univariable logistic regression, Spearman’s correlation analysis, the least absolute shrinkage and selection operator (LASSO) algorithm, and the Boruta algorithm. Six ML models were constructed, with performance evaluated using area under the curve (AUC), calibration curves, and decision curve analyses (DCA) to identify the optimal model. Five-fold and ten-fold cross-validation were used to assess the optimal model’s generalizability, and its interpretability was evaluated via Shapley Additive exPlanations (SHAP) analysis.ResultsA total of 347 patients were included, comprising a training set (n=193, January 2016–December 2024), validation set (n=84, January 2016–December 2024), and test set (n=70, January 2025–July 2025). From 164 initial features, 7 key features were retained through a four-step screening. The Random Forest (RF) model outperformed other models, achieving AUCs of 0.99 (95% CI: 0.97–1.00) in the training set, 0.88 (95% CI: 0.80–0.95) in the validation set, and 0.97 (95% CI: 0.94–1.00) in the test set. Calibration curve and decision curve analyses confirmed its strong clinical utility. Five-fold cross-validation for the RF model showed fold-specific AUCs of 0.82–0.92, with a mean AUC of 0.87 (95% CI: 0.84–0.90). Ten-fold cross-validation showed fold-specific AUCs of 0.80–0.99, with a mean AUC of 0.88 (95% CI: 0.83–0.93). SHAP analysis revealed five novel spatial anatomical features (such as Sagittal plane-posterior spatial anatomical structure index, Coronal plane-Left anatomical structure interval) were negatively associated with PSM risk, while the number of positive biopsy cores and clinical tumor stage were positively associations.ConclusionsMulti-dimensional fusion data combined with ML models improves PSM prediction accuracy in RARP. The RF model, with excellent performance and interpretability, shows promise for preoperative PSM risk stratification, facilitates optimized clinical decision-making, and supports personalized treatment discussions during preoperative planning, but requires prospective and external validation before clinical implementation.

Annual burned areas in the United States have increased twofold during the past decades. With more large fires resulting in more emissions of fine particulate matter, an accurate prediction of fire emissions is critical for quantifying the impacts of fires on air quality, human health, and climate. This study aims to construct a machine learning (ML) model with game-theory interpretation to predict monthly fire emissions over the contiguous US and to understand the controlling factors of fire emissions. By comparing the predicted fire PM2.5 emissions from the interpretable ML model with the Global Fire Emissions Database (GFED) observations and predictions from process-based models in the Fire Modeling Intercomparison Project (FireMIP), the ML model is also used to diagnose the process-based models to inform future development. Results show promising performance for the ML model, Community Land Model (CLM), and Joint UK Land Environment Simulator-Interactive Fire And Emission Algorithm For Natural Environments (JULES-INFERNO) in reproducing the spatial distributions, seasonality, and interannual variability of fire emissions over CONUS. Regional analysis shows that only the ML model and CLM simulate the realistic interannual variability of fire emissions for most of the subregions (r > 0.95 for ML and r = 0.14 ~ 0.70 for CLM), except for Mediterranean California, where all the models perform poorly (r = 0.74 for ML and r < 0.30 for the FireMIP models). Regarding seasonality, most models capture the peak emission in July over western US. However, all models except for the ML model fail to reproduce the bimodal peaks in July and October over Mediterranean California, which may be explained by the coarse spatial resolutions of the processed-based models or atmospheric forcing data or limitations in model parameterizations for capturing the effects of Santa Ana winds on fire activity. Furthermore, most models struggle to capture the spring peak in emissions in southeastern US, probably due to underrepresentation of human effects and the influences of winter dryness on fires in the models. As for extreme events, both the ML model and CLM successfully reproduce the frequency map of extreme emission occurrence but overestimate the number of months with extremely large fire emissions. Comparing the fire PM2.5 emissions from the interpretable ML model with process-based fire models highlights their strengths and uncertainties for regional analysis and prediction and provides useful insights on future directions for model improvements.

In this study, we introduce a novel approach that integrates interpretability techniques from both traditional machine learning (ML) and deep neural networks (DNN) to quantify feature importance using global and local interpretation methods. Our method bridges the gap between interpretable ML models and powerful deep learning (DL) architectures, providing comprehensive insights into the key drivers behind model predictions, especially in detecting outliers within medical data. We applied this method to analyze COVID-19 pandemic data from 2020, yielding intriguing insights. We used a dataset consisting of individuals who were tested for COVID-19 during the early stages of the pandemic in 2020. The dataset included self-reported symptoms and test results from a wide demographic, and our goal was to identify the most important symptoms that could help predict COVID-19 infection accurately. By applying interpretability techniques to both machine learning and deep learning models, we aimed to improve understanding of symptomatology and enhance early detection of COVID-19 cases. Notably, even though less than 1% of our cohort reported having a sore throat, this symptom emerged as a significant indicator of active COVID-19 infection, appearing 7 out of 9 times in the top four most important features across all methodologies. This suggests its potential as an early symptom marker. Studies have shown that individuals reporting sore throat may have a compromised immune system, where antibody generation is not functioning correctly. This aligns with our data, which indicates that 5% of patients with sore throats required hospitalization. Our analysis also revealed a concerning trend of diminished immune response post-COVID infection, increasing the likelihood of severe cases requiring hospitalization. This finding underscores the importance of monitoring patients post-recovery for potential complications and tailoring medical interventions accordingly. Our study also raises critical questions about the efficacy of COVID-19 vaccines in individuals presenting with sore throat as a symptom. The results suggest that booster shots might be necessary for this population to ensure adequate immunity, given the observed immune response patterns. The proposed method not only enhances our understanding of COVID-19 symptomatology but also demonstrates its broader utility in medical outlier detection. This research contributes valuable insights to ongoing efforts in creating interpretable models for COVID-19 management and vaccine optimization strategies. By leveraging feature importance and interpretability, these models empower physicians, healthcare workers, and researchers to understand complex relationships within medical data, facilitating more informed decision-making for patient care and public health initiatives.

To develop an interpretable machine learning (ML) model for predicting surgical outcomes in renal hilar tumors and propose a hilar-specific anatomical nephrometry scoring system. A total of 414 patients with renal hilar tumors who underwent robot-assisted partial nephrectomy (RAPN) were included in this study, comprising 304 patients from the First Affiliated Hospital of Nanchang University and 110 patients from the Second Affiliated Hospital of Nanchang University, which served as the external validation cohort. To identify predictors of trifecta achievement, we used least absolute shrinkage and selection operator regression and the Boruta algorithm, followed by multivariate logistic regression to identify independent factors. Five ML models were developed and evaluated using receiver operating characteristic curves, calibration plots, decision curve analysis, and precision-recall curves. The generalizability of the model was further validated in external cohort. Finally, we used SHapley Additive exPlanations (SHAP) to interpret the contribution of each predictor and enhance the model's explainability. Furthermore, based on anatomical features identified through logistic regression, we developed a modified nephrometry scoring system and compared its risk stratification performance with the traditional R.E.N.A.L. (i.e., Radius, Exophytic or endophytic, Nearness, Anterior or posterior, and Location) scoring system. Among the 304 patients in the primary cohort, 168 achieved trifecta outcomes. Eight variables were incorporated into the predictive model, with logistic regression model ultimately being selected as the optimal predictive model. It showed robust predictive performance in internal and external validation. SHAP methods identified surgeon, classification of hilar tumor, and radius as the three most significant predictive variables. Compared with the traditional R.E.N.A.L. score, the modified R.E.N.A.L. score demonstrated superior stratification ability for operation time, change in serum creatinine, change in estimated glomerular filtration rate, and trifecta achievement. The interpretable ML model accurately predicts trifecta in RAPN for hilar tumors. The modified R.E.N.A.L. score provides refined anatomical stratification and facilitates individualized surgical planning.

To establish an interpretable and non-invasive machine learning (ML) model using clinicoradiological predictors and magnetic resonance imaging (MRI) radiomics features to predict the consistency of pituitary macroadenomas (PMAs) preoperatively. Total 350 patients with PMA (272 from Xinqiao Hospital of Army Medical University and 78 from Daping Hospital of Army Medical University) were stratified and randomly divided into training and test cohorts in a 7:3 ratio. The tumor consistency was classified as soft or firm. Clinicoradiological predictors were examined utilizing univariate and multivariate regression analyses. Radiomics features were selected employing the minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) algorithms. Logistic regression (LR) and random forest (RF) classifiers were applied to construct the models. Receiver operating characteristic (ROC) curves and decision curve analyses (DCA) were performed to compare and validate the predictive capacities of the models. A comparative study of the area under the curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE) was performed. The Shapley additive explanation (SHAP) was applied to investigate the optimal model's interpretability. The combined model predicted the PMAs' consistency more effectively than the clinicoradiological and radiomics models. Specifically, the LR-combined model displayed optimal prediction performance (test cohort: AUC = 0.913; ACC = 0.840). The SHAP-based explanation of the LR-combined model suggests that the wavelet-transformed and Laplacian of Gaussian (LoG) filter features extracted from T2WI and CE-T1WI occupy a dominant position. Meanwhile, the skewness of the original first-order features extracted from T2WI (T2WI_original_first-order_Skewness) demonstrated the most substantial contribution. An interpretable machine learning model incorporating clinicoradiological predictors and multiparametric MRI (mpMRI)-based radiomics features may predict PMAs consistency, enabling tailored and precise therapies for patients with PMA.