Hu Sihan, Guo Xiaochuan, Wang Xiaobao, Jin Zixiang, Zhou Chenyang, Tu Lang, Shi Zhoulong, Ao Weiyi, Zhang Xin, Zheng Jay, Zhang Xuezhi, Ye Hui
Department of Integrated Traditional Chinese and Western Medicine, Peking University First Hospital, Institute of Integrated Traditional Chinese and Western Medicine, Peking University, Beijing, China.
Department of Respiratory Diseases, The First Affiliated Hospital of Henan University of Chinese Medicine, Zhengzhou, China.
Front Med (Lausanne). 2025 Aug 26;12:1633890. doi: 10.3389/fmed.2025.1633890. eCollection 2025.
Chronic digestive system diseases (CDSD) pose a major health challenge worldwide, significantly increasing morbidity and mortality rates. The frailty index is crucial for assessing patient prognosis. To address the need for proactive healthcare, we developed a multi-timepoint frailty prediction model.
This study collected data from 565 patients with CDSD, including their frailty assessments at 3 and 6 years of follow-up. Utilizing the Multi-Gate Mixture-of-Experts (MMoE) framework, we built and evaluated five models: Tab Transformer, Convolutional Neural Network (CNN), Deep Neural Network (DNN), Extreme Gradient Boosting (XGBoost) and Random Forest (RF). We comprehensively compared the predictive capabilities of these models on both validation and test sets.
The MMoE framework consistently outperforms single models in predicting both 3-year and 6-year frailty indices across most metrics. Specifically, for 3-year predictions, the single model achieves an accuracy of 0.9801 (95% CI: 0.963-0.990) on the train set and 0.5487 (95% CI: 0.457-0.637) on the test set, while the MMoE model reaches 0.956 (95% CI: 0.933-0.971) and 0.982 (95% CI: 0.938-0.995), respectively. The RF model demonstrated perfect performance, with Micro-AUC values of 1.000 in both training and test sets for both 3-year and 6-year intervals, leading other models in terms of accuracy, precision, recall, F1 score. The Tab Transformer model achieved high Micro-AUC values across all prediction intervals, with values of 0.997 and 0.995 in the training set for 3-year and 6-year predictions, respectively, and corresponding test set values of 0.999 and 0.987.
This MMoE-based approach can predict frailty at key time points, offering insights into frailty progression and aiding clinical decision making. Integrating this AI model into CDSD management can promote early interventions and personalized treatment plans.
慢性消化系统疾病（CDSD）在全球范围内构成了重大的健康挑战，显著增加了发病率和死亡率。衰弱指数对于评估患者预后至关重要。为了满足积极医疗保健的需求，我们开发了一种多时间点衰弱预测模型。
本研究收集了565例CDSD患者的数据，包括他们在随访3年和6年时的衰弱评估。利用多门专家混合（MMoE）框架，我们构建并评估了五个模型：表格变换器（Tab Transformer）、卷积神经网络（CNN）、深度神经网络（DNN）、极端梯度提升（XGBoost）和随机森林（RF）。我们在验证集和测试集上全面比较了这些模型的预测能力。
在大多数指标上，MMoE框架在预测3年和6年衰弱指数方面始终优于单一模型。具体而言，对于3年预测，单一模型在训练集上的准确率为0.9801（95%置信区间：0.963 - 0.990），在测试集上为0.5487（95%置信区间：0.457 - 0.637），而MMoE模型分别达到0.956（95%置信区间：0.933 - 0.971）和0.982（95%置信区间：0.938 - 0.995）。RF模型表现出色，在3年和6年间隔的训练集和测试集中，微平均曲线下面积（Micro - AUC）值均为1.000，在准确性、精确性、召回率、F1分数方面领先于其他模型。表格变换器模型在所有预测间隔中均获得了较高的微平均曲线下面积值，在3年和6年预测的训练集中分别为0.997和0.995，在测试集中相应的值为0.999和0.987。
这种基于MMoE的方法可以在关键时间点预测衰弱，深入了解衰弱进展并辅助临床决策。将这种人工智能模型整合到CDSD管理中可以促进早期干预和个性化治疗方案。
