DeMMO: Longitudinal and Cross-Disease Modelling of Digital Mobility Outcomes via Multi-Task Learning DeMMO：通过多任务学习对数字移动性结果进行纵向和跨疾病建模

TL;DR

• DeMMO is an interpretable multi-task learning framework for longitudinal, multi-disease, and multi-outcome modelling of digital mobility outcomes (DMOs) from wearable sensors
• It introduces an automatic cross-disease and cross-outcome relation-learning mechanism that discovers signed relationships directly from longitudinal coefficient matrices, enabling selective information sharing even when disease cohorts have no overlapping participants
• The framework combines temporal regularization with stable and visit-specific feature selection to model how multivariate DMO-clinical outcome relationships evolve jointly across diseases
• Evaluated on the large-scale Mobilise-D dataset with 24 harmonized DMOs across five visits and multiple mobility-limiting conditions, DeMMO outperforms nine strong linear, longitudinal, and deep-regression baselines
• Stability selection identifies reliable longitudinal DMO patterns with potential for clinical validation and disease monitoring

Why It Matters

This work addresses a critical gap in digital health research: most wearable-based studies analyze a single disease at a single timepoint, missing the opportunity to understand how mobility biomarkers relate to clinical outcomes across diseases and over time. For AI practitioners working in healthcare and time-series modeling, DeMMO demonstrates how multi-task learning with structured regularization can enable knowledge transfer across disjoint cohorts—a common real-world constraint in clinical data. The interpretable nature of the framework also aligns with the growing demand for explainable AI in medical applications.

Technical Details

• Longitudinal DMO coefficient matrix: Each disease-outcome objective is represented by a coefficient matrix capturing the relationship between wearable-derived mobility features and clinical outcomes across visits, enabling explicit modeling of temporal evolution
• Cross-disease/cross-outcome relation learning: The core innovation is an automatic mechanism that learns signed (positive/negative) relations directly from longitudinal mappings, allowing selective information sharing between diseases and outcomes without requiring paired participants across cohorts
• Temporal regularization + feature selection: DeMMO combines temporal smoothness constraints with stable (shared across visits) and visit-specific feature selection, balancing consistency and adaptability in longitudinal patterns
• Evaluation on Mobilise-D: Tested on a multicentre dataset with 24 harmonized real-world DMOs over five visits across multiple mobility-limiting conditions, compared against nine baselines spanning linear models, longitudinal methods, and deep regression architectures
• Interpretability via stability selection: The framework produces clinically interpretable longitudinal DMO patterns ranked by selection stability, supporting downstream validation and monitoring use cases

Industry Insight

• The ability to learn cross-disease relationships from disjoint cohorts is highly relevant to real-world healthcare AI, where patient-level data sharing across studies is often restricted by privacy regulations—this approach could accelerate biomarker discovery without requiring centralized data pooling
• As wearable sensor adoption grows in clinical trials and remote patient monitoring, frameworks like DeMMO that can jointly model multiple outcomes across conditions will become increasingly valuable for building comprehensive digital phenotyping pipelines
• The emphasis on interpretability through stability selection reflects a broader industry shift toward clinically deployable AI; practitioners should consider how model transparency and feature-level insights can facilitate regulatory approval and clinician adoption in digital health applications

TL;DR

• 提出DeMMO框架，实现纵向、多疾病、多结果的可解释学习，突破现有研究单疾病单时间点分析的局限
• 核心创新是自动跨疾病和跨结果关系学习机制，直接从纵向映射中学习有符号关系，无需配对参与者即可实现选择性信息共享
• 在Mobilise-D大规模多中心数据集上验证，涵盖24个协调真实世界DMOs、5次访问和多种行动限制条件
• 相比9个线性、纵向和深度回归基线方法，DeMMO取得最佳整体和特定结果预测性能
• 稳定性选择识别出可靠的纵向DMO模式，支持后续临床验证和疾病监测

为什么值得看

DeMMO解决了数字移动结果(DMO)研究中多疾病联合建模的关键空白，为可穿戴设备辅助的慢性病纵向监测提供了可解释的方法论框架。其跨疾病信息共享机制在缺乏配对参与者数据时仍能实现有效学习，对医疗AI落地具有重要参考价值。

技术解析

DeMMO将每个疾病-结果目标表示为纵向DMO系数矩阵，结合时间正则化与稳定且访问特定的特征选择。其核心机制是自动跨疾病和跨结果关系学习，直接从纵向映射中学习有符号关系，实现选择性信息共享。实验基于Mobilise-D数据集，包含24个协调真实世界DMOs、5次访问和多种行动限制条件。相比9个强基线（线性、纵向、深度回归），DeMMO在整体和特定结果预测上均取得显著改进。稳定性选择进一步识别可靠纵向DMO模式。

行业启示

可穿戴传感器与多任务学习的结合为慢性病数字化监测开辟了新路径，推动AI从单病种分析向多病种联合建模演进。跨疾病信息共享机制为数据稀缺场景下的模型训练提供了可行方案，可推广至其他医疗领域。模型的可解释性设计使其输出更易被临床验证，加速AI医疗应用从研究到实践的转化。