开发并验证一种以单一切面超声心动图视频作为输入的多任务深度学习模型，以预测肥厚型心肌病（HCM）及其伴发征象。

回顾性纳入2016年1月至2025年5月在武汉大学人民医院就诊的HCM及对照组（非HCM组）受试者377例（2组总人数按照约1∶2进行配对），收集每例受试者静息状态下经胸标准胸骨旁长轴切面（PLAX）二维灰阶超声视频数据。全部数据集以随机分层抽样的方法按4∶1的比例划分为训练集及测试集。基于单一切面视频训练多任务深度学习模型，深度学习模型并输出预测HCM、左心室流出道梗阻（LVOTO）及二尖瓣收缩期前向运动（SAM）征的概率值。在训练集中进行模型的4-折交叉验证和最终版本模型的开发，并在测试集中评估最终版本模型3个任务的受试者操作特征曲线下面积（ROC-AUC）及精确率-召回率曲线下面积（PR-AUC）。

377例受试者中HCM组患者117例（179个视频），对照组260例（273个视频），共收集452个视频（其中HCM阳性视频数179个，LVOTO阳性视频数111个，SAM征阳性视频数91个）。测试集中，最终版本深度学习模型预测HCM的ROC-AUC及PR-AUC分别为0.987［95%置信区间（CI）：0.966~0.999］、0.981（95%CI：0.945~1.000）；预测LVOTO的ROC-AUC及PR-AUC分别为0.899（95%CI：0.769~0.978）、0.789（95%CI：0.585~0.925）；预测SAM征的ROC-AUC及PR-AUC分别为0.980（95%CI：0.943~1.000）、0.940（95%CI：0.837~1.000）。

基于单一切面超声心动图视频的多任务深度学习模型能同时准确识别HCM及其伴发征象LVOTO和SAM征，该模型有作为实时决策支持工具和降低HCM漏诊的潜力。

This study aimed to develop and validate a multi-task deep learning model that uses a single-plane echocardiographic video as input to predict hypertrophic cardiomyopathy (HCM) and its associated signs.

In this retrospective study, 377 subjects were included, comprising patients with HCM and a control group (non-HCM group) who visited Renmin Hospital of Wuhan University between January 2016 and May 2025 (the two groups were matched at a ratio of approximately 1∶2). Standard transthoracic two-dimensional gray-scale echocardiographic video data of the parasternal long-axis view (PLAX) at rest were collected for each subject. The entire dataset was divided into a training set and a test set at a 4∶1 ratio using random stratified sampling. A multi-task deep learning model was trained on single-plane video data and simultaneously outputs probability estimates for predicting HCM, LVOTO, and SAM. On the training set, the model underwent 4-fold cross-validation, and the final version of the deep learning model was developed. The performance of the final model was evaluated on the test set using the area under the receiver operating characteristic curve (ROC-AUC) and the area under the precision-recall curve (PR-AUC) for the three tasks.

Among the 377 subjects, 117 had HCM and 260 were in the control group. A total of 452 videos were collected (including 179 positive HCM videos, 111 positive LVOTO videos, and 91 positive SAM videos). In the test set, the ROC-AUC and PR-AUC of the final version of the deep learning model for predicting HCM were 0.987 [95%(confidence interval, CI): 0.966-0.999] and 0.981 (95%CI: 0.945-1.000), respectively; for predicting LVOTO, they were 0.899 (95%CI: 0.769-0.978) and 0.789 (95%CI: 0.585-0.925); and for predicting SAM, they were 0.980 (95%CI: 0.943-1.000) and 0.940 (95%CI: 0.837-1.000).

This multi-task deep learning model based on single-view echocardiographic video can accurately and simultaneously identify HCM and its associated signs (LVOTO and SAM), demonstrating its potential as a real-time decision support tool and helping reduce missed diagnoses of HCM.