BMC medical research methodology · 2026
Quality and performance of machine learning versus logistic regression for predicting IVIG resistance in Kawasaki disease: a PROBAST+AI systematic comparison.
Сравнительное исследованиеСтатьяМетаанализНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Zhang Jiaying, Wang Difan, Dong Jinfeng, He Ying, Liu Ying, You Tingjiao, Li Jing, Li Lizhi, Wu Xiaodan, Tang Qiuyu, Ma Shurong, Liu Panpan, Lv Haitao, Huang Hongbiao
Аннотация
BACKGROUND: This study aimed to systematically compare the predictive performance and methodological quality of logistic regression (LR) and machine learning (ML) models for intravenous immunoglobulin (IVIG) resistance in Kawasaki disease (KD) using the PROBAST + AI framework.
METHODS: We searched PubMed, Embase, and Web of Science to identify studies on prediction models for IVIG resistance in KD published between January 1, 2006, and July 31, 2025. We assessed methodological rigour, risk of bias, and applicability using PROBAST + AI. A meta-analysis was performed using random-effects models with logit-transformed area under the receiver operating characteristic curve (AUC) values. Subgroup, sensitivity, and publication bias analyses were additionally conducted.
RESULTS: We identified 52 eligible studies (40 LR and 12 ML). In external validation, pooled AUCs were similar between ML and LR models (0.76 [95% CI 0.64-0.86] vs. 0.75 [95% CI 0.68-0.81]). In internal validation, ML showed a slightly higher pooled AUC than LR (0.86 [95% CI 0.78-0.92] vs. 0.76 [95% CI 0.72-0.79]), although no statistically significant differences were observed. All studies were judged to be at high risk of bias, mainly due to retrospective single-centre designs, inadequate handling of missing data and continuous predictors, and poor reporting of calibration and clinical utility. No study reported sample size calculations.
CONCLUSIONS: Given the limited external validation and substantial heterogeneity across studies, ML does not consistently outperform LR in predicting IVIG resistance in KD. Future studies should prioritise rigorous external validation and adherence to TRIPOD + AI and PROBAST + AI.
METHODS: We searched PubMed, Embase, and Web of Science to identify studies on prediction models for IVIG resistance in KD published between January 1, 2006, and July 31, 2025. We assessed methodological rigour, risk of bias, and applicability using PROBAST + AI. A meta-analysis was performed using random-effects models with logit-transformed area under the receiver operating characteristic curve (AUC) values. Subgroup, sensitivity, and publication bias analyses were additionally conducted.
RESULTS: We identified 52 eligible studies (40 LR and 12 ML). In external validation, pooled AUCs were similar between ML and LR models (0.76 [95% CI 0.64-0.86] vs. 0.75 [95% CI 0.68-0.81]). In internal validation, ML showed a slightly higher pooled AUC than LR (0.86 [95% CI 0.78-0.92] vs. 0.76 [95% CI 0.72-0.79]), although no statistically significant differences were observed. All studies were judged to be at high risk of bias, mainly due to retrospective single-centre designs, inadequate handling of missing data and continuous predictors, and poor reporting of calibration and clinical utility. No study reported sample size calculations.
CONCLUSIONS: Given the limited external validation and substantial heterogeneity across studies, ML does not consistently outperform LR in predicting IVIG resistance in KD. Future studies should prioritise rigorous external validation and adherence to TRIPOD + AI and PROBAST + AI.