Xiao Hu

Health time series (HTS) provide a fundamental representation of dynamic health processes, spanning organ-level physiological signals, patient-level clinical trajectories, and population-level epidemic dynamics. Despite their shared temporal nature, existing surveys often study these categories separately, leaving their methodological connections underdeveloped. In this survey, we provide a unified review of deep learning for HTS, organized around a scarcity-driven methodological development trajectory. Specifically, we examine how label scarcity motivates task-specific supervised learning with domain-aware temporal modeling; how sample scarcity motivates foundation model development and generative augmentation; and how information scarcity motivates multi-modal learning by incorporating complementary clinical information. Finally, we summarize representative healthcare applications and open research directions. By connecting data categories, scarcity challenges, and methodological paradigms, this survey provides a structured roadmap for understanding and developing deep learning methods for HTS.

Objective: Complex predictive models are increasingly proposed as substitutes for laboratory measurements, yet they are typically assessed with aggregate accuracy metrics that reveal little about downstream clinical impact. We developed an evaluation framework that assesses a prediction model by its utility for guiding decisions under realistic resource constraints, demonstrated through non-invasive lactate estimation in the intensive care unit (ICU).Methods: The framework assumes the model is used to derive a resource allocation policy that selects which ICU patients to re-measure lactate and which to rely on the model's estimate, under a pre-specified budget for invasive blood draws. Policies are evaluated by average clinical cost of errors, area under the receiver operating characteristic curve (AUROC) for hyperlactatemia classification at the clinical threshold, and regression mean squared error (MSE). Beyond random allocation and prediction-ranked or threshold-distance heuristics, we propose Predictive Risk-Informed Sampling (PRIS), which estimates each patient's false-negative and false-positive risk with separate classifiers and ranks by a cost-weighted score, prioritizing draws for cases most vulnerable to harmful misclassification. On MIMIC-III (788 ICU stays) with a pretrained lactate predictor, we compared PRIS against heuristic baselines across budget levels.Results: PRIS significantly reduced average clinical cost versus random allocation at low-to-moderate budgets. All rules improve as measured values replace predictions, but PRIS reaches consequential errors sooner, resolving 22 of 61 misclassifications by 50 draws versus 14 for the prediction-ranked baseline, and achieving the highest classification performance at every budget. Because it ranks by error probability rather than predicted magnitude, PRIS led in clinical cost and classification but was least effective for regression error: the preferred rule depended on the evaluation metric, a distinction that aggregate accuracy would not reveal. At the budget of the routine draw-all-abnormal rule, prediction-ranked identified none of the 34 missed hyperlactatemia cases, whereas PRIS caught 14.Conclusion: Aggregate accuracy does not capture a model's clinical usefulness; evaluating decision utility under a measurement budget is a better alternative. As a triage layer on an existing lactate predictor, PRIS directs lab resources toward cases where inaccurate predictions carry the highest risk of harmful clinical consequences.

Was not provided for this article.

OBJECTIVES: Sepsis is a heterogeneous syndrome resulting from a complex interaction between the infectious agent and the host response. In adults with sepsis, longitudinal temperature measurements have been used to identify phenotypes with distinct clinical characteristics, immune profiles, and outcomes. Our objective was to develop and validate temperature trajectory-based phenotypes in critically ill children with sepsis. We hypothesized that the temperature trajectory-based phenotypes would be independently associated with clinical outcomes. DESIGN: Retrospective observational cohort study. SETTING: A total of 13 PICUs in the United States from January 1, 2012, to January 1, 2018. PATIENTS: Children (age 0–18 y) with community-acquired sepsis meeting Phoenix sepsis criteria admitted to a PICU directly or via an emergency department, who survived to 24 hours and had greater than or equal to 4 temperature measurements in the first 24 hours of admission were included. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: A total of 11,566 patients met all criteria for study inclusion. A total of 2327 patients (20%) developed the primary outcome of persistent multiple organ dysfunction syndrome (MODS) assessed at day 7 of admission. We used group-based trajectory modeling to identify temperature trajectory-based phenotypes. Three groups were identified: “hyperthermic” (35%), “normothermic” (59%), and “hypothermic” (6%), which were reproducible and had similar outcomes on a validation set. The “hypothermic” trajectory-based phenotype group had the highest odds of persistent MODS odds ratio (OR), 5 (95% CI, 4–6), and ICU mortality OR, 11 (95% CI, 9–15). In an exploratory analysis, hypothermia (< 35°C) at the time of PICU admission was also associated with persistent MODS and ICU mortality. The association between “hypothermic” trajectory-based phenotype and poor outcome remained after adjusting for hypothermia at admission, severity of illness, age, and technology dependence. CONCLUSIONS: We developed and validated temperature trajectory-based phenotypes of pediatric sepsis in a large multicenter cohort. The “hypothermic” trajectory-based phenotype was independently associated with worse clinical outcomes.

Clinical decision-making (CDM) is central to real-world clinical workflows, where clinicians infer diagnoses, select treatments, or anticipate future health outcomes under incomplete evidence. LLMs are increasingly used to support these decisions due to strong language capabilities, broad biomedical knowledge, and efficiency, yet the reliability of LLMs on real-world clinical decision tasks remains insufficiently understood. To evaluate CDM models, especially LLM-based models, an ideal and practical medical decision benchmark should be constructed via an automated yet reliable pipeline to ensure both scale and quality. Moreover, the grounding of a CDM benchmark in real patient EHRs can better support evaluation on practical CDM tasks that require substantive biomedical knowledge and clinical inference. To fill the gaps, we introduce EHRBench, an automated and reliable EHR-grounded benchmark for evaluating LLM-based clinical decision-making at scale. To ensure scalability and reliability, EHRBench is constructed through an EHR-LLM-KB (knowledge-base) interaction pipeline. For efficiency, we use a specialized LLM to automatically convert encounter-level EHR trajectories into structured templates and deterministically instantiate the templates into QA items. In parallel, we apply systematic KB-based verification and enrichment to filter hallucinated or ambiguous relations and to improve reliability. Using this pipeline, we construct nearly 1M (960,067) QA items spanning three core inference-required clinical decision tasks: diagnosis, treatment, and prognosis. We benchmark more than 30 representative LLMs on EHRBench and provide detailed analyses of performance and robustness. The results show consistent capability trends across settings, further validating the reliability of EHRBench and highlighting actionable gaps toward clinically reliable LLM systems.

BACKGROUND Chronic disease management increasingly depends on monitoring symptoms and health behaviors outside clinical settings, yet most mobile health (mHealth) systems treat behavioral data capture and intervention delivery as separate functions. Platforms that collect high-resolution sensor or ecological momentary assessment (EMA) data often provide limited tools for clinician-authored intervention content, whereas intervention applications commonly embed content and decision logic directly in application code. This coupling slows iteration, limits personalization, and creates dependence on software developers. OBJECTIVE This paper presents CanBeWell, a disease-agnostic telehealth platform that integrates multimodal behavioral data collection with a clinician-operated, code-free intervention authoring and delivery workflow. CanBeWell consists of a patient-facing mobile application, a backend application programming interface (API) and data layer, and a clinician-facing dashboard. We describe the platform's system architecture, design rationale, implementation for cancer survivorship care, and formative stakeholder evaluation. METHODS CanBeWell was developed as a platform-oriented system informed by prior work in remote patient monitoring, mHealth sensing, and adaptive intervention theory. The platform was evaluated through formative semi-structured interviews with cancer patients (n=12) and their caregivers (n=9). [CX3.1]All study participants reviewed the mobile application prototype and provided feedback during in-person interviews on its usability, perceived value, and suggested improvements. Participants also rated the platform's overall usability on a 10-point scale. Additionally, clinicians (n = 6) were provided access to both the patient-facing mobile application and the clinician dashboard to evaluate the platform's usability, functionality, clinical relevance, and intervention workflow. Qualitative data were analyzed using thematic analysis by two independent coders RESULTS Twenty-one participants (patients and caregivers) completed the demographic survey (mean age 49.4 years; 61.9% female; 57.1% non-Hispanic White). Four themes emerged based on patient and caregiver interviews: ease of use, interest in the application, valued features, and overall user-friendliness[CX4.1]. The mean usability rating across patients and caregivers was 9 out of 10. Patients and caregivers valued the phased intervention structure, baseline activity tracking, and incremental goal progression. Clinicians responded positively to the no-code authoring workflow, noting that the ability to review behavioral data, draft content, and deploy approved phases without developer involvement could reduce technical barriers and accelerate intervention refinement[CX5.1]. The most common requested improvement by clinicians was stronger notification support to sustain engagement over longer deployments. CONCLUSIONS CanBeWell extends remote patient monitoring infrastructure by adding a clinician-governed intervention authoring and delivery layer that closes the loop between behavioral sensing and patient-facing support. Formative validation in a cancer survivorship context suggests that the platform is usable and clinically meaningful to patients, caregivers, and clinicians. By separating intervention content from application code, the platform enables clinicians to create and update personalized interventions without extensive programming support, reducing engineering bottlenecks and facilitating scalable implementation across chronic disease settings.

Objective: Accurate classification of physiological signals in real-world deployments is challenged by sensor noise, motion artifacts, and distribution shifts between training and deployment data. Inference-time augmentation (ITA), which applies augmentations during inference rather than retraining, offers a simple, model-agnostic mechanism to improve robustness. However, ITA application to physiological signals has remained narrow in scope, relying on limited augmentation methods with fixed, unoptimized parameters. This work proposes a unified ITA framework to address that gap. Approach: The framework incorporates 13 augmentation methods spanning time-domain, amplitude-domain, frequency-domain, and artifact-injection transformations, with hyperparameters systematically optimized via Bayesian optimization. We evaluate the framework on atrial fibrillation (AF) detection from 30-second photoplethysmography (PPG) signals using two deep learning architectures, GPT-PPG (in two sizes) and ResNet, across five datasets comprising more than 400 patients and ~9,800 hours of PPG recording. Two evaluation strategies are assessed: standard ITA applied to all inputs, and selective ITA applied to initially positive predictions. Main results: Standard ITA consistently improved the area under the receiver operating characteristic curve (AUROC, up to 8.5% for GPT-PPG and 0.7% for ResNet) and the area under the precision-recall curve (AUPRC, up to 10.6% for GPT-PPG and 0.8% for ResNet) across all model–dataset combinations. Selective ITA further reduced average FPR by up to 4.4% (GPT-PPG) and 1.3% (ResNet) on non-AF PPG datasets. Significance: These findings establish ITA as a practical, model-agnostic approach for improving the reliability of PPG-based AF classification in deployment settings where retraining is not feasible, with broader applicability to physiological signal analysis.