Hospital Records Show Missing Clinical Data Influences Machine Learning Predictions

A retrospective evaluation of routine cardiology records indicates that the choice of missing-data handling strategy modestly influences machine-learning performance, while algorithmic architecture accounts for most predictive variation. Researchers Gönül Kara and Leyla Özgür Polat examined how incomplete electronic hospital encounters affect the identification of documented cardiovascular cases. Rather than treating missingness purely as an administrative flaw to be erased through imputation, the study analyzed whether the absence of clinical measurements itself conveys informative signals regarding clinical workflows and diagnostic documentation.

Standard machine-learning investigations of cardiovascular disorders frequently depend on curated benchmark datasets that are fully populated and unrepresentative of standard clinical settings. In everyday hospital environments, incomplete documentation occurs routinely, yet the methodological handling of missing values is rarely analyzed as an explicit experimental factor. To address this discrepancy, the researchers investigated seven distinct missing-data strategies using routine cardiology encounter records to assess whether missingness reflects operational documentation behaviors rather than random oversights.

The study cohort comprised 11,706 encounter records collected between 2019 and 2024 from the cardiology department of a private hospital in Denizli, Türkiye. Investigators coded these encounters into 18 candidate predictor variables alongside a primary outcome defined by a documented cardiac diagnosis. Following the predefined exclusion of encounters with six or more unrecorded variables, the analytic cohort contained 7,680 records. Within this final dataset, 58.1 percent of patient encounters featured the documented cardiac diagnosis outcome.

To assess methodological variability, the team benchmarked seven operational approaches across identical data splits. These strategies included single imputation, native algorithmic handling of missing values, the addition of an explicit unknown category, complete case analysis, multiple imputation, iterative tree imputation, and imputation augmented with explicit missingness indicators. Complete case analysis substantially constrained sample size, restricting the evaluable cohort to 2,465 fully observed records across the designated clinical variables.

Twelve classifier architectures were systematically tuned utilizing a five-by-five repeated cross-validation design. The investigators evaluated predictive utility and model reliability using the area under the receiver operating characteristic curve, calibration measures, and discrete class metrics accompanied by bootstrap confidence intervals and DeLong tests. Interpretability was assessed via SHAP values, and the primary comparisons were reinforced through 18 sensitivity analyses alongside four supplementary experimental evaluations.

Across all evaluated scenarios, tree-based ensemble models consistently outperformed competing approaches, with Random Forest and CatBoost ranking either first or second in every scenario. CatBoost achieved its highest discriminative performance under the unknown category approach, yielding an area under the receiver operating characteristic curve of 0.822 with a 95 percent confidence interval spanning 0.813 to 0.831. By comparison, CatBoost achieved an area under the receiver operating characteristic curve of 0.812 under single imputation.

The inclusion of an unknown category and the use of missingness indicators both exceeded single imputation in pooled discriminative accuracy by margins ranging between 0.004 and 0.023 across every tested algorithm. These improvements met statistical significance under Holm-corrected DeLong tests across 10 and 11 of the 11 compared models, respectively, and across 4 and 7 models under Nadeau–Bengio corrections. Conversely, neither multiple imputation nor iterative tree imputation outperformed single imputation. When investigators introduced simulated missingness generated independently of diagnostic outcomes, no strategy outperformed single imputation.

Variance decomposition revealed that the choice of machine-learning algorithm explained 72.5 percent of the variance in cross-validation performance, whereas the missing-data strategy explained only 4.1 percent. Complete case analysis generated the lowest observed area under the receiver operating characteristic curve at 0.795 with CatBoost; however, single imputation applied to that same subset of 2,465 records yielded equivalent results with zero of 11 models showing significant differences. Furthermore, an indicator tracking unmeasured blood pressure—denoting records lacking both a measurement and documented hypertension—emerged among the three most influential predictive features in 23 of 25 cross-validation folds.

The authors concluded that among leading machine-learning architectures, selecting between imputation techniques altered discriminative performance by at most 0.014, an impact smaller than the performance disparity between distinct algorithm families. The apparent performance penalty observed under complete case analysis was consistent with a reduced and differently structured sample rather than the deletion mechanism alone. Because missingness indicators captured diagnostic and administrative documentation practices, the researchers emphasized that such signals reflect recording behavior rather than direct physiological evidence.

Subject of Research: Machine-learning performance and missing-data strategies in routine hospital cardiovascular records

Article Title: Missing-data governance in real-world cardiovascular case identification from routine hospital records: effects on machine-learning performance and interpretability

Article References: Kara, G., & Polat, L. Ö. (2026). Missing-data governance in real-world cardiovascular case identification from routine hospital records: effects on machine-learning performance and interpretability. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03976-2

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03976-2

Keywords: cardiology, machine learning, missing data, CatBoost, random forest, AUROC, hospital records, electronic health records, data governance, imputation methods, clinical informatics, cardiovascular disease, predictive modeling, SHAP values, cross-validation, Denizli, Turkey