Systematic assessment of mixed imputation methods and explainable machine learning

Abstract

The convergence of artificial intelligence and precision oncology is frequently hampered by the quality of real-world clinical data, particularly the pervasive challenge of missing values. This opinion review critically appraises the methodology and evidentiary framework of the study, which proposes a hybrid imputation architecture, HDI-MF-Gower, integrated with an extra trees classifier and Shaply Additive exPlanation interpretability for predicting survival outcomes following curative gastrectomy. We deconstruct the pivotal assumptions and potential sensitivities of their adaptive weighted similarity initialization. This design is engineered to provide a “warm start” aligned with the underlying data structure for iterative imputation, theoretically mitigating the risks of distributional distortion associated with simplistic initialization strategies. However, a primary boundary of the current evidence lies in the validation hierarchy; the reported validation relies predominantly on random splitting within a single-center cohort, lacking the robustness of temporal extrapolation or genuine external validation. Furthermore, statistical comparisons suggest that the performance differences between the proposed model and several robust ensemble baselines are not consistently distinguishable, making it difficult to attribute performance gains solely to the specific choice of the learner. We conclude that future research must construct a more rigorous evidence chain within multicenter and multimodal frameworks. Crucially, adherence to transparent reporting of a multivariable prediction model for individual prognosis or diagnosis + artificial intelligence guidelines - specifically regarding missing data mechanisms, sensitivity analyses, calibration and net benefit assessments, and the availability of reproducible materials - is essential to substantiate generalizable clinical utility.

Key Words: Gastric cancer; Machine learning; Survival prediction; Missing data imputation; Extra trees

Core Tip: This opinion review evaluates the study, which introduces a hybrid imputation framework, HDI-MF-Gower, combined with an explainable extra trees classifier for predicting postoperative survival in gastric cancer. The study’s innovative approach addresses the challenges of missing clinical data by adapting the iterative imputation method MissForest. Although the model demonstrates clinical transparency and methodological rigor, its marginal improvement over existing ensemble methods and lack of external validation highlight areas for further research in multimodal data integration and multi-institutional validation.

INTRODUCTION

Gastric cancer represents a formidable challenge in global oncology and constitutes a critical public health issue. According to the latest Global Cancer Observatory statistics, gastric cancer ranks fifth in global incidence and fourth in mortality[1]. The burden of this disease is disproportionately distributed geographically, with the highest incidence clusters observed in East Asia, particularly in China, Japan, and the Republic of Korea. China alone accounts for a substantial proportion of the global caseload, suggesting the imperative for developing localized, data-driven risk stratification and decision-support tools specifically tailored to this population[2].

The clinical management of gastric cancer has undergone a significant paradigm shift, evolving from surgery-centric approaches to multimodal strategies involving D2 lymphadenectomy, perioperative chemotherapy, and targeted biological agents[3]. Despite these therapeutic advances, the postoperative survival trajectories of patients undergoing radical gastrectomy remain characterized by marked heterogeneity. Two patients with identical tumor-node-metastasis (TNM) staging may experience starkly contrasting outcomes: One may suffer early recurrence within months, while the other achieves long-term disease-free survival[4]. This discrepancy underscores a fundamental limitation of our current prognostic framework: The TNM system, predicated primarily on anatomical stratification, captures a limited dimensionality of information and is therefore insufficient to fully characterize the complex interplay between tumor biology, host status, and perioperative factors[5,6].

To bridge this prognostic gap, the oncology community is increasingly pivoting toward artificial intelligence (AI) and machine learning[7-9]. These computational paradigms possess the capacity to process high-dimensional datasets and discern non-linear interactions and complex patterns that elude traditional Cox proportional hazards models. However, the translation of machine learning from in silico experimentation to clinical bedside application faces two pervasive obstacles: Data quality and model interpretability[7,10-14].

Real-world clinical data is inherently messy. Unlike the pristine, curated datasets often utilized in computer science competitions, electronic health records are plagued by missing values[15,16]. Laboratory tests are omitted, records are inconsistent, and patient follow-up is often fragmented[17]. Traditional statistical approaches to handling missingness - such as complete case analysis, which discards any record with a missing value - can introduce selection bias and alter sample representativeness if the missingness is not completely random. This not only diminishes statistical power but, more critically, compromises the validity of the findings. Conversely, simplistic imputation methods (e.g., mean imputation) may underestimate variable variance and uncertainty, thereby impairing model calibration and extrapolation robustness[18-21].

It is against this backdrop that the manuscript by Lü et al[22] serves as a representative case study for discussing evidentiary standards regarding missing data handling, validation frameworks, and interpretability in clinical prediction models. The authors attempt to construct a workflow integrating the hybrid imputation algorithm HDI-MF-Gower, the extra trees classifier, and Shaply Additive exPlanation (SHAP) analysis[11]. This opinion review aims to provide a granular analysis of this work. We will scrutinize the critical methodological choices regarding their imputation strategy, its underlying assumptions, and potential sources of sensitivity; assess the robustness of their validation strategies and generalization evidence against existing methodological consensus; benchmark their reporting against the newly released transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) + AI 2024 guidelines; and discuss the broader implications for external validation, deployment, and reproducible practices in future multimodal surgical data science[23-25].

DECONSTRUCTING THE HDI-MF-GOWER ARCHITECTURE

To evaluate its incremental value and define its applicability boundaries against conventional imputation benchmarks - such as multivariate imputation by chained equations (MICE), MissForest, and K-nearest neighbors (KNN) - this section delineates the critical design choices inherent to the workflow as detailed. Furthermore, we identify potential sources of sensitivity that may compromise the robustness of the study’s conclusions (Figure 1).

Figure 1 Schematic representation of the HDI-MF-Gower hybrid imputation framework. A: Warm start initialization: Raw data containing missing values (white blocks) are processed using an adaptive weighted Gower distance. Unlike standard mean imputation, this step assigns dynamic weights (w_k) to variables based on their variance and missing rates, ensuring that high-information donors drive the initial “guess” (light blue blocks); B: Iterative optimization: The initialized matrix enters a MissForest loop. For each variable, a random forest regressor/classifier captures non-linear interactions and updates the missing entries. The cycle repeats until the difference (δ) between consecutive iterations falls below the convergence threshold (γ), yielding the final imputed dataset.

Theoretical framework of missing data

Prior to evaluating the specific imputation strategy, it is pertinent to clarify the prevailing theoretical frameworks governing missing data mechanisms. In clinical research, missingness is typically conceptualized through Rubin’s classification: Missing completely at random (MCAR), missing at random, and missing not at random (MNAR)[16,17].

The authors report conducting Little’s MCAR test, noting a lack of statistical evidence significantly deviating from the MCAR assumption[21]. It is crucial to emphasize, however, that a non-significant result in this context does not definitively establish that the data are MCAR; rather, it merely indicates a failure to reject the null hypothesis. Furthermore, this test often lacks sufficient statistical power in the presence of complex missingness patterns or limited sample sizes[18,19]. Consequently, a more robust methodological approach involves not only reporting missingness rates and patterns but also explicitly discussing the plausibility of missing at random and potential MNAR scenarios. For instance, if patients with more severe disease or poorer prognoses are less likely to undergo certain follow-up assessments, then the absence of those data points itself carries prognostic information - a clear case of an MNAR mechanism. Ideally, this should be accompanied by sensitivity analyses to quantify how varying imputation strategies might influence model performance and interpretative conclusions.

Initialization based on adaptive Gower distance

Many iterative imputation algorithms, such as MissForest and MICE, require an initial imputation step to prime the iterative process. If these initial values diverge significantly from the data’s latent distribution, they may compromise convergence rates and the stability of the final imputed results, potentially amplifying bias risks in scenarios characterized by high missingness or complex correlation structures. To address this challenge, Lü et al[22] propose a similarity-based initialization step - a “warm start” - leveraging weighted Gower distance. Gower distance serves as a similarity metric tailored for mixed-type data, facilitating sample matching across datasets comprising both continuous and categorical variables. They assign variable-specific weights based on each feature’s information content and missingness rate, so that more informative and more complete variables receive higher influence in the distance calculation. This phase culminates in matching each incomplete sample to similar individuals (nearest neighbors), from which initial imputation values are generated. Compared to initialization via global mean or mode, this similarity-based “warm start” is intuitively more aligned with individual heterogeneity and may attenuate coarse distortions of marginal variable distributions. Nevertheless, the net benefit vs potential bias warrants quantification through rigorous sensitivity analyses, and the stability of subsequent interpretative conclusions under varying imputation settings requires further verification[26,27].

Iterative optimization based on MissForest

Following similarity-based initialization, the authors employ MissForest for iterative imputation within their workflow. MissForest is a non-parametric method that reframes imputation as a supervised learning problem. The algorithm treats each variable with missing entries as a distinct prediction task: During the iterative process, a random forest is trained for each incomplete variable, utilizing the remaining variables as predictors to update the missing entries. This cycle repeats until the difference between imputation matrices from consecutive iterations falls below a preset threshold γ or a maximum number of iterations is reached.

The selection of random forest over linear regression models - common in standard MICE implementations - reflects a deliberate intent to leverage its capacity for fitting non-linear relationships and high-order interactions. In the context of perioperative and oncological variables for gastric cancer, threshold effects or conditional dependencies are clinically commonplace[18-20]. For instance, risk associations for certain continuous markers may exhibit non-linearity, while the prognostic contribution of staging or metastasis-related variables may vary depending on clinical contexts such as tumor location. Random forest inherently captures these complex structures, thereby reducing the reliance on the explicit pre-specification of interaction terms[28].

Overall, HDI-MF-Gower can be conceptualized as a “warm-start” modification of MissForest: By generating initial imputed values that are structurally closer to the data via similarity matching, it theoretically influences convergence speed and imputation error. The authors report that, under the experimental missingness rates presented (e.g., 5%-20%), this method achieves lower normalized root mean squared error and proportion of falsely classified compared to KNN, MICE, and standard MissForest. It is important to note, however, that the generalizability of such comparative results remains contingent upon implementation details, including the missingness mechanism, variable preprocessing (scaling/encoding), weight and neighbor settings, and stopping criteria. If intended as a precursor to clinical modeling, it is advisable to further quantify the impact of these choices on downstream model performance and interpretative stability through sensitivity analyses and uncertainty reporting[18-20,28].

CLINICAL MODELING AND INTERPRETABILITY

Following imputation, the authors implemented a feature selection protocol, designating the intersection of least absolute shrinkage and selection operator and Boruta outputs as the candidate feature set. This intersectional strategy adopts a conservative stance, designed to mitigate feature instability and potential redundancy inherent in high-dimensional, small-sample scenarios, although its efficacy remains contingent upon specific implementation and data structures. Akin to random forests, extra trees (extremely randomized trees) typically construct trees using the entire sample without bootstrapping, while introducing an enhanced stochastic splitting mechanism - randomly selecting cut-points rather than optimizing them - to inject randomness and thereby reduce model variance. In settings characterized by limited sample size (n = 304) and relatively high dimensionality, the risk of overfitting warrants particular vigilance. While some tree-based models may exhibit sensitivity to noise - especially if hyperparameter tuning, feature selection, and evaluation pipelines are not strictly nested - the aggressive randomization of extra trees functions as a form of implicit regularization, potentially alleviating overfitting and enhancing generalization[24,25]. The authors report an area under the curve (AUC) of 0.853 in their validation/test setting, outperforming comparative models[29,30]. However, discrimination metrics alone are insufficient to substantiate clinical utility. If the model is intended for risk stratification, we advocate for the concurrent reporting of uncertainty intervals, calibration performance (e.g., calibration curves or Brier scores), and potential clinical net benefit (decision curve analysis), substantiated by more rigorous internal and external/temporal validation[31-33]. In clinical prediction landscapes, beyond discriminatory performance, the interpretability, communicability, and auditability of model outputs often determine their integration into real-world decision-making workflows. The authors employ SHAP, which decomposes predictions into feature contributions under axiomatic guarantees, ensuring an additive representation where the sum of contributions aligns with the model output. The SHAP summary plot provides a global overview of feature attribution: TNM staging remains a dominant predictor, while variables associated with nutritional/inflammatory status, tumor burden, and metastatic potential - such as albumin, CEA, and lymphovascular invasion - also exhibit substantial contributions. The compatibility of these findings with established clinical knowledge lends a degree of face validity to the model. However, this biological plausibility does not preclude the risk of structural artifacts introduced by imputation and feature engineering[34,35]. A more robust evidentiary chain should necessitate reporting on interpretative stability - specifically, whether SHAP rankings and effect directions remain consistent across different imputation schemes, resampling iterations, or temporal splits. Furthermore, it must be noted that SHAP reflects model attribution rather than causal effects; in the presence of strong feature correlation, attribution allocation may shift, further underscoring the critical importance of stability testing[36-39].

CRITICAL APPRAISAL: THE “REVIEWER DETECTIVE” AUDIT

Based on the information disclosed in the main text, this section synthesizes the study’s primary contributions and evaluates the specific areas requiring reinforcement regarding methodological increments, evidentiary validation, and reproducibility (Table 1).

Table 1 Alignment of the study with transparent reporting of a multivariable prediction model for individual prognosis or diagnosis + artificial intelligence (2024) reporting domains and key statistical checks.

Evaluation domain	TRIPOD + AI item	Study implementation	Critical commentary
Data integrity	Item 11: Missing data handling	Strong: Used HDI-MF-Gower hybrid imputation; adaptive initialization	The hybrid design is intended to improve initialization for iterative imputation; however, missingness assumptions and MNAR-oriented sensitivity analyses are not reported
Validation rigor	Item 10/22: Validation strategy	Weak: Random split (7:3) within single-center cohort (type 2a)	A random split within a single-center cohort may inflate performance estimates; temporal and/or geographical external validation is not shown
Discrimination	Item 13a: Model performance	Moderate: AUC = 0.853 (extra trees)	Statistically superior to KNN/support vector machine (P < 0.05), but not significant vs XGBoost (P = 0.098) or RF (P = 0.196). Gain may be data-driven rather than model-driven
Calibration	Item 13b: Calibration plot	Gap: Not explicitly detailed in primary comparison	Essential for clinical utility. High AUC does not guarantee accurate risk probability estimation
Interpretability	Item 22: Model explanation	Strong: SHAP summary and dependence plots used	Identified clinically relevant markers (albumin, lymphovascular invasion), providing surface validity. Needs stability testing across folds
Reproducibility	Item 25: Availability of code/model	Gap: Online tool mentioned but access/archiving unclear	Full adherence requires public repository (e.g., GitHub) and clear version control for clinical deployment Availability is not clearly documented (no persistent link, repository, or archived version); providing code/model cards, versioning, and a persistent access route would strengthen reproducibility

TRIPOD: Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis; AI: Artificial intelligence; MNAR: Missing not at random; AUC: Area under the curve; KNN: K-nearest neighbors; RF: Random forest; SHAP: Shaply Additive exPlanation.

Methodological deconstruction

Structurally, HDI-MF-Gower represents an orchestration and procedural adaptation of existing components - specifically, similarity metrics and tree-based imputation - rather than the introduction of a de novo statistical framework. While one might argue this lacks theoretical novelty, such engineering adaptations often generate meaningful gains in performance and usability within the pragmatic constraints of clinical prediction. In applied medical informatics, where data are often characterized by small sample sizes, missingness, and measurement heterogeneity, novel but complex architectures do not inherently outperform robust, modular combinations[18-20]. Consequently, the targeted modification of established components (e.g., Gower distance combined with tree-based imputation) to address mixed-type clinical data holds distinct engineering value. The authors report improvements in metrics such as normalized root mean squared error under their experimental missingness settings, suggesting that the “similarity-based warm start + iterative tree imputation” combination yields synergistic gains[24]. To render this increment fully verifiable, we recommend a more comprehensive disclosure of comparative conditions - including missingness mechanism settings, parameter/iteration stopping criteria, repetition counts, and uncertainty intervals - alongside sensitivity analyses to test the stability of this advantage across varying missingness structures and preprocessing configurations[40].

Validation hierarchy and generalizability

The use of the term “independent validation set” in the title and abstract may inadvertently mislead readers into interpreting this as external validation; however, the methodology describes a random split (e.g., 7:3) within a single-center cohort, which constitutes internal validation[41]. Training and validation samples in this setting likely share site-specific idiosyncrasies regarding clinical pathways and measurement processes (e.g., surgical teams, pathological interpretation, and perioperative management), thereby tending to yield optimistic performance estimates that may degrade during cross-institutional application[23-25,40,42]. A more robust validation trajectory necessitates a clear distinction between internal, temporal, and external validation. For the internal tier, we recommend bootstrapping or nested cross-validation with the reporting of optimism-corrected performance[29]. If extrapolation is claimed, temporal validation based on enrollment time is the minimum requirement; genuine external validation must stem from distinct centers, regions, or data acquisition workflows[43]. Furthermore, beyond discrimination metrics, clinical application claims should be supported by calibration assessments (calibration curves, Brier scores) and potential clinical net benefit (decision curve analysis)[31-33]. The manuscript’s discussion acknowledges the limitations of the single-center retrospective design and the need for external validation; this transparency helps define the boundaries of applicability but also signals that the current evidence grade requires elevation through a more rigorous validation system.

Statistical benchmarking and attribution

The final version of the manuscript supplements the statistical comparison of AUCs across models. Results indicate that while the difference between extra trees and weaker baselines (KNN, support vector machine) is statistically significant (P < 0.05), the differences against XGBoost (P = 0.098) and random forest (P = 0.196) do not reach significance. Given the multiple pairwise comparisons, it is advisable to report effect sizes and confidence intervals for these differences and clarify whether corrections for multiplicity were applied, enabling readers to distinguish between statistical significance and potential clinical relevance. Collectively, these results suggest that on this specific dataset and feature set, the discriminative performance of various robust ensemble methods converges. Consequently, current evidence is insufficient to attribute performance advantages solely to the choice of a specific learner. It is plausible that upstream data preparation steps, such as imputation and feature selection, contribute significantly to overall performance. To clarify the sources of performance gain, we suggest conducting further ablation or replacement experiments and supplementing these with calibration and decision curve analyses, which are more directly relevant to clinical utility[32,39,44]. Finally, achieving statistical significance in these comparisons is not a substitute for rigorous validation; even a model with significant internal gains requires confirmation on external datasets to establish generalizability.

BENCHMARKING AGAINST TRIPOD+AI (2024) GUIDELINES

Benchmarked against the core tenets of the TRIPOD+AI (2024) statement, the manuscript demonstrates commendable transparency in several domains, notably in the detailed description of the missing data imputation workflow and the reporting of model implementation and software environment specifications[23,45]. Furthermore, the utilization of SHAP provides valuable interpretative cues at both global and local levels. Conversely, critical gaps determining the evidence grade persist, primarily clustering around validation hierarchy and reproducibility/accessibility[24,25,40]. The current validation strategy - resembling an internal random split within a single-center cohort - lacks the evidentiary support of temporal validation or genuine external validation[29]. If clinical transferability is claimed, supplementary assessments relevant to application, such as calibration and clinical net benefit, are requisite. Another verifiable deficit concerns the public availability of the online tool: The text fails to provide a persistent access link, archiving details, or usage instructions, thereby leaving deployment and auditability unproven. While these gaps do not negate the engineering integration value of the work, they directly compromise its evidence grade and cross-institutional transferability[25,46]. In line with the TRIPOD+AI reproducibility domain (item 25), making the code and predictive models publicly available is necessary to substantiate the study’s findings and enable external validation.

FUTURE OUTLOOK: BEYOND CLINICAL TABULAR DATA

While HDI-MF-Gower offers a systematic imputation pathway for tabular clinical variables, prognostication relying exclusively on structured data - such as age, TNM staging, and laboratory indices - is inherently constrained by the dimensionality of the information. The reported AUC of approximately 0.85 indicates a competent discriminative capacity; however, elevating performance and clinical utility likely necessitates the integration of data streams that more directly characterize tumor biology, alongside a rigorous evaluation of calibration and clinical net benefit[24,25,31].

The trajectory of multimodal fusion lies in the joint modeling of tabular clinical variables with radiomics (e.g., texture and heterogeneity features from preoperative computed tomography) and pathomics (e.g., quantitative characterization of the tumor microenvironment and immune infiltration on whole slide images) to address the limitations of single-modality characterization of tumor aggressiveness[47-52]. Existing reviews indicate that in many settings, intermediate fusion - integrating distinct modal features within a shared latent space - frequently yields superior performance compared to unimodal or simple late fusion strategies[53-55]. However, given the substantial variability across datasets and evaluation protocols, generalizability must be confirmed through strict external validation[56-58]. Future iterations might consider positioning HDI-MF-Gower as the imputation module specifically for the clinical data stream within a fused architecture, while simultaneously evaluating the impact of diverse missingness structures and fusion strategies on performance and interpretative stability[38,58]. Notably, data integrity across different modalities remains a challenge - methods like HDI-MF-Gower will need to ensure reliable imputation of missing clinical features to fully realize the benefits of multimodal fusion.

Under the constraints of privacy and data sovereignty, federated learning provides a viable pathway for multicenter training: Models are trained locally at multiple institutions, aggregating parameter updates to absorb cross-center heterogeneity without sharing raw patient data[59]. This approach helps mitigate the limitations of single-center training and creates conditions for constructing an evidence chain that approximates external validation. Nevertheless, clinical transferability must still be verified using pre-defined, cross-institutional independent test sets - incorporating temporal extrapolation and geographically distinct cohorts where necessary - while concurrently reporting application-relevant metrics such as calibration and decision curves[60]. Moreover, although federated learning mitigates the need for data pooling, it does not automatically resolve site-specific missing data problems; careful imputation and data quality control at each participating center remain necessary to preserve model fidelity.

CONCLUSION

The study by Lü et al[22] presents an integrated workflow centering on missing data handling, tree-based modeling, and interpretable outputs, representing an incremental advancement at the level of methodology and engineering implementation. By introducing a weighted similarity “warm start” within a mixed-variable space, HDI-MF-Gower offers a verifiable implementation pathway for iterative tree imputation; the authors report lower error metrics compared to several baseline methods under their experimental missingness settings. The combination of extra trees and SHAP provides visual cues for feature contribution, with certain high-contribution features aligning with established clinical knowledge. However, the stability of these interpretative conclusions requires verification across different imputation settings, resampling iterations, and temporal splits.

Concurrently, the primary boundary of the current evidence lies in the validation hierarchy: The random splitting employed is closer to internal validation, lacking the support of temporal extrapolation and genuine external validation. Furthermore, DeLong comparisons suggest that the AUC differences between the proposed model and several robust ensemble methods may not be statistically distinguishable. Regarding the online tool, the deployment statement currently lacks critical information such as a persistent access link, versioning, and usage instructions; consequently, public availability and reproducibility remain unproven, limiting immediate translation and peer review.

Looking toward the future research agenda, we recommend systematically reporting missingness rates and patterns while discussing missing data mechanisms. Future work should conduct sensitivity analyses for imputation strategies and stability tests for interpretation; adopt more rigorous internal validation while clearly distinguishing between internal, temporal, and external validation hierarchies; supplement claims of clinical application with calibration and decision curves; and provide code, model cards, and versioned tool archiving. These steps are essential to synchronously elevate the grade of evidence and clinical transferability. In summary, only through thorough real-world validation and careful calibration can these AI-driven models achieve the trust and utility needed for clinical practice.