医学
机器学习
原发性醛固酮增多症
队列
人工智能
支持向量机
人工神经网络
内科学
诊断准确性
队列研究
前馈神经网络
接收机工作特性
平均绝对误差
计算机科学
作者
Wenyu Zhang,Christina Pamporaki,René Jäkel,Georgiana Constantinescu,Mirko Peitzsch,Manuel Schulze,Jun Yang,Andrea Horvath,R. Müller-Pfefferkorn,Martín Reincke,Felix Beuschlein,Carmina Teresa Fuß,Stefanie Hahner,Graeme Eisenhofer
标识
DOI:10.1093/ejendo/lvag096.223
摘要
Abstract Background Commonly used screening tests for primary aldosteronism provide suboptimal diagnostic accuracy. Machine learning (ML) models that utilize steroid profiles offer an alternative approach but require improvement. Methods Training and validation of ML models—using a panel of 13 plasma steroids, potassium, and renin—involved three patient cohorts, a total of 1380 patients, tested after medication washout. Models were initially trained on one cohort (n = 352) to facilitate comparisons of a previous model also trained on that cohort; the second cohort (n = 849) was used for external validation. Selected best models were also tested with and without medication washout. Models were subsequently trained on the combined cohorts and externally validated on the third cohort (n = 179). Diagnostic performance of best models was then compared to performance of the aldosterone-to-renin ratio (ARR). Findings Integration of plasma steroids with potassium using a feedforward neural network (FNN) model provided improved diagnostic performance over a previously established support vector machine model according to areas under ROC curves (AUCs) of 0.940 (CI:0.918-0.963) and 0.917 (CI:0.893-0.937, P = .0103). Addition of renin negligibly improved model performance beyond the improvement with potassium. The FNN and other renin-independent models showed similar accuracy before and after antihypertensive medication washout, whereas renin-dependent models exhibited poorer accuracy without medication washout. Paired comparisons of combined-dataset-models with the ARR showed three further optimized renin-independent models outperformed the ARR for screening according to AUCs of 0.948 to 0.954 vs 0.839 for the ARR. Those models, which utilize five selected steroids, minimize need for medication washout and at cut-offs for optimal 90% to 95% diagnostic sensitivity reduce false-positives by 53% to 72% to more effectively screen for PA than with the ARR. Interpretation The newly developed ML models demonstrate enhanced diagnostic performance compared to the previously established support vector machine model and the ARR. There was no difference in the diagnostic performance of renin-independent compared to renin-dependent models, though the former offer the benefit of screening patients in whom medication washout is not feasible. Funding eutsche Forschungsgemeinschaft.
科研通智能强力驱动
Strongly Powered by AbleSci AI