Research on postpartum depression risk prediction and intervention based on nursing-driven multimodal data fusion
护理驱动下多模态数据融合的产后抑郁风险预测及干预研究

摘要

目的 构建基于多模态数据与可解释AI的产后抑郁（PPD）动态预测与分层干预系统，并验证其在护理场景中的早期识别与干预效果。方法 采用类随机对照设计，将产后6小时内符合条件的300例产妇采用类随机对照分为观察组与对照组，各150例。观察组接受AI预测+分层干预，对照组接受常规护理。于产后7、14、28、42天评估EPDS、HRV、睡眠效率及母婴互动异常，并比较风险识别时效、EPDS变化、高危缓解率及护理负荷。结果 观察组风险识别更快（1.21±0.68天 vs 4.73±1.92天，P<0.05），EPDS下降更明显（−5.62±3.14 vs −2.13±2.87，P<0.05），高危缓解率更高（92.0% vs 64.0%，P<0.05）。护理负荷同步下降（随访耗时减少34.7%，随访人数与压力评分显著降低）。模型AUC为0.864，F1为0.79，可解释AI提高了护理人员接受度。结论 该多模态可解释AI体系可提升PPD早期识别与干预效果，改善情绪症状并提高护理效率，具有良好应用前景。

Abstract

Objective To develop a multimodal, explainable AI–based system for dynamic prediction and stratified intervention of postpartum depression (PPD) and to evaluate its effectiveness in early identification and intervention within nursing settings.
Methods In a quasi-randomized controlled study, 300 postpartum women within 6 hours after delivery were assigned to an observation group or a control group (150 each). The observation group received AI-assisted prediction and stratified care, while the control group received routine nursing. EPDS, HRV, sleep efficiency, and maternal–infant interaction abnormalities were assessed on postpartum days 7, 14, 28, and 42. Risk identification timeliness, EPDS changes, high-risk remission rates, and nursing workload were compared.
Results The observation group showed faster risk identification (1.21±0.68 vs. 4.73±1.92 days, P<0.05), greater EPDS reduction (−5.62±3.14 vs. −2.13±2.87, P<0.05), and a higher remission rate (92.0% vs. 64.0%, P<0.05). Nursing workload decreased substantially, with reduced follow-up time, follow-up volume, and stress levels. The model achieved an AUC of 0.864 and an F1 score of 0.79, and explainable outputs improved nurses ' acceptance.
Conclusion   This multimodal explainable AI system enhances early PPD risk detection and intervention effectiveness, improves maternal emotional outcomes, and reduces nursing workload, demonstrating strong potential for clinical application.