Technical Summary: Personalized Digital Health Modeling with Adaptive Support Users

Problem Statement

Personalized machine learning is critical in digital health due to substantial physiological and behavioral heterogeneity among individuals. However, effective personalization is hindered by the scarcity, noise, and fragmentation of user-specific data. When individual data are insufficient to train stable models, external data are necessary, yet incorporating them without distorting personal signals remains a challenge.

Existing approaches typically rely on population pretraining followed by fine-tuning or pooling data only from users deemed "similar" to the target individual. These strategies suffer from two main limitations:

1. Biased Transfer: Treating all external data uniformly (population pretraining) or restricting transfer to a small neighborhood of similar users ignores the potential value of dissimilar users.
2. Missed Opportunities: Dissimilar users, whose data may encode informative contrasts or counterexamples, are often excluded. This leads to limited coverage of population-level variation and prevents the model from learning what the target user is not.

Directly including data from dissimilar users without adaptation can degrade accuracy by introducing misleading patterns. There is a need for a training method that can adaptively leverage external data to transfer beneficial information while suppressing detrimental signals.

Methodology

The authors propose a unified personalization framework that trains a personal model using adaptive support users, a group comprising both similar and dissimilar individuals. The core innovation is the formulation of personalization as a single optimization objective that jointly updates model parameters and learned adaptive weights.

1. Support User Initialization

The process begins by computing similarity scores $s(u, j)$ between a target user $u$ and all external users $j$. This is achieved using cosine similarity between user embeddings derived from an auxiliary neural network trained on pooled population data. Unlike hard clustering, all users are retained as potential support users, acknowledging that both highly similar and highly dissimilar users may provide valuable information.

2. Unified Optimization Objective

The framework minimizes a composite loss function (Equation 2) consisting of three terms:

• Personal Loss ($L$): Direct supervision from the target user's own data ($D_u$), ensuring specificity to individual patterns.
• Similar-User Knowledge Transfer: A weighted sum of losses from similar users ($S(u)$). The contribution of each similar user is modulated by a fixed similarity score $s(u, j)$ and a learned adaptive weight $\alpha_j$. This term transfers predictive structure from aligned patterns.
• Dissimilar-User Penalty ($R$): A contrastive regularization term from dissimilar users ($D(u)$). This term uses a margin-based loss to prevent the personal model from adopting misleading patterns found in dissimilar users, effectively structuring the representation space to separate the target user from those with contrasting profiles.

The objective is subject to a simplex constraint ($\sum \alpha_j = 1, \alpha_j \geq 0$), ensuring normalized influence.

3. Iterative Optimization Algorithm

The framework employs an alternating optimization procedure:

• Step 1 (Update Model Parameters): With adaptive weights $\alpha_j$ fixed, the model parameters (shared encoder $\phi_{\theta_u}$ and personalized head $g_{\theta_u}$) are updated via backpropagation. Gradients from the personal loss, transfer term, and penalty term shape the encoder and head in complementary ways.
• Step 2 (Update Adaptive Weights): With model parameters fixed, the weights $\alpha_j$ are updated based on the predictive utility of each support user. A cost function $c_j$ is computed, balancing the prediction error on similar users against the representational overlap with dissimilar users. Weights are then normalized via softmax, assigning higher influence to users that provide beneficial generalization or informative contrast.

Key Contributions

1. Redefinition of Support Users: The paper introduces the concept of support users comprising both similar and dissimilar individuals, arguing that dissimilar users provide meaningful complementary information (counterfactual structure) often overlooked in prior work.
2. Unified Adaptive Framework: A novel framework that integrates heterogeneous support user data by learning adaptive contributions. This allows the model to selectively leverage helpful signals and suppress detrimental ones through a principled optimization process.
3. Empirical Validation and Interpretability: Extensive validation across six healthcare tasks on four real-world datasets demonstrates the method's efficacy. Furthermore, the learned adaptive weights provide interpretability, offering actionable insights for targeted data selection.

Experimental Results

The framework was evaluated on four real-world digital health datasets: Loneliness, Affect, CGMacro (glucose response), and Globem (sleep forecasting). Performance was compared against six baselines, including population models, purely personalized models, merged models, transfer-based models, cluster-based models, and fixed similarity-weighted models.

• Performance Gains: The proposed method consistently outperformed baselines. On large-scale datasets (e.g., Globem), it achieved up to 10% lower RMSE. In low-data settings, improvements were even more pronounced, reaching approximately 25% lower RMSE.
• Data Efficiency: Experiments varying the percentage of support users showed that models trained on selected subsets (top 10% most similar and top 10% most dissimilar) often outperformed those trained on the full population, validating the utility of the "support user" concept.
• Architecture Sensitivity: Simpler MLP encoders outperformed Transformer encoders in small-data regimes (Loneliness, Affect), while the performance gap narrowed as data availability increased (Globem).
• Ablation Studies: Removing either the similar-user transfer term or the dissimilar-user penalty term led to significant performance degradation in many tasks, confirming the necessity of both components. Similarly, using fixed weights instead of adaptive ones resulted in lower performance, particularly in heterogeneous populations.

Significance and Claims

The paper claims that its primary significance lies in shifting the paradigm of personalized digital health from excluding dissimilar users to adaptively integrating them. By treating personalization as a unified objective that balances personal loss, positive transfer, and contrastive regularization, the method addresses the fundamental challenge of data scarcity and noise.

The authors emphasize that the learned adaptive weights serve a dual purpose:

1. Performance: They dynamically optimize the trade-off between generalization and specificity, leading to superior predictive accuracy.
2. Interpretability: The weights provide a transparent mechanism for understanding which external users contribute most to a specific individual's model. This offers practical guidance for data collection strategies, suggesting that future data gathering should prioritize not only the target user but also those support users with high adaptive weights.

The work concludes that this approach enables more accurate monitoring and recommendations in digital health by effectively leveraging the full spectrum of population diversity, rather than just the most similar subset.