Comment to the article - Machine-learning model for postoperative atrial fibrillation - Letter to the Editor

Sep 29, 2026

Upload by Finn

Letter responding to a study evaluating a machine-learning model for predicting postoperative atrial fibrillation after isolated CABG. The authors question reliability and clinical utility due to insufficient sample size and low events-per-variable, artificially balanced outcome prevalence, a limited hold-out test set with no external validation, and unclear algorithm choice and lack of explainability. They also note inconsistencies between reported performance metrics and the confusion matrix, and emphasize cautious interpretation and the need for robust future reporting.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Letter to the Editor

Insufficient sample size and events-per-variable (EPV)

Artificially balanced outcome prevalence

Limited test set and absence of external validation

Choice of algorithms and interpretability

Summary and recommendations

Turkish Journal of Thoracic and Cardiovascular Surgery 2025;33(4):590-591 Letter to the Edıtor / Editöre Mektup Comment to the article: Machine-learning model for postoperative atrial fibrillation Makaleye yorum: Postoperatif atriyal fibrilasyon için makine-öğrenimi modeli Fatih Yiğit􀀬, Taylan Adademir􀀬, Kaan Kırali􀀬 Department of Cardiovascular Surgery, Koşuyolu High-Specialization Training and Research Hospital, İstanbul, Türkiye We read with great interest the study by Akbulutet al. on machine-learning (ML) prediction of postoperative atrial fibrillation (POAF) after isolated CABG. [1] Although the integration of artificial intelligence into peri-operative care is welcome, several methodological issues may overstate the model’s reliability and clinical value . Insufficient sample size and events-per-variable (EPV) Although the initial model considered 91 candidate predictors, only ≈15 features appear to have been retained after Boruta-based selection . With 50 POAF events in 100 patients, the resulting EPV is approximately 3.3-well below the level required for model stability. A model with 15 predictors and an anticipated outcome incidence of ~22% would require at least 550-700 participants to ensure reliable parameter estimation, prevent overfitting, and achieve a shrinkage factor ≥0.9 . [2,3] The current sample therefore falls far short of the minimum sample size required for internally valid model development under modern reporting standards . Artificially balanced outcome prevalence The cohort was sub-sampled to a 50/50 POAF-non-POAF split, creating an artificial prevalence of 0.50. All reported performance metrics in Table 3-such as sensitivity, specificity, precision, F1 score, accuracy, and Cohen’s κ-reflect this engineered balance of 50% rather than the true clinical incidence of POAF, which is ≈ 22 % . Recent simulation work shows that such balancing inflates apparent accuracy and produces severe mis-calibration when the model is applied to real-world data . [4] Limited test set and absence of external validation Only 20 patients comprised the hold-out test set; misclassifying a single case alters accuracy by five percentage points . No geographic or temporal validation was reported, contrary to TRIPOD-AI reporting guidance . [2] Choice of algorithms and interpretability The core model utilizes a Probabilistic Data Association (PDA) classifier-originally designed for radar and sonar tracking, not for clinical binary classification . Moreover, no model explainability method (e.g. , SHAP) was provided, despite transparent interpretation being essential to clinical applicability and trust. Moreover, the confusion matrix in Figure 4 is inconsistent with the sensitivity, specificity and precision values reported in the text (e.g. , TP=10, FP=1 yield sensitivity=1.00 and specificity ≈ 0.90, not vice versa), indicating a reporting error and raising doubt about reliability of the performance estimates . Corresponding author: Fatih Yiğit. E-mail: [faithygt90@hotmail.com](faithygt90@hotmail.com) [Doi: 10.5606/tgkdc.dergisi.2025.28291](Doi: 10.5606/tgkdc.dergisi.2025.28291) [Received:](Received: July 02)[ July 02](Received: July 02) , 2025 Accepted: September 24, 2025 Published online: October 20, 2025 Cite this article as: Yiğit F, Adademir T, Kırali K. Comment to the article: Machine-learning model for postoperative atrial fibrillation. Turk Gogus Kalp Dama 2025;33(4):590-591 . doi: 10.5606/tgkdc.dergisi.2025.28291 . ©2025 All right reserved by the Turkish Society of Cardiovascular Surgery. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes ([http://creativecommons.org/licenses/by-nc/4.0/](http://creativecommons.org/licenses/by-nc/4.0/)). 590 [https://tgkdc.dergisi

FAQs

What methodological concerns are raised about the machine-learning model?

The letter highlights insufficient sample size and low events-per-variable, artificially balanced outcome prevalence, limited testing with no external validation, and issues affecting reporting and interpretation of performance.

How does outcome balancing affect the reported metrics?

Sub-sampling created an artificial 50/50 prevalence, which the letter says inflates apparent accuracy and leads to mis-calibration when applied to real-world postoperative atrial fibrillation incidence.

Why is the reported model performance questioned?

The hold-out test set contains only 20 patients, so misclassifying one case notably changes accuracy, and the confusion matrix is reported as inconsistent with values stated in the text.

Sep 29, 2026

Upload by Finn