Health-LLM: Large Language Models for Health Prediction via Wearable Sensor Data

• Health-LLM evaluates 12 LLMs across 10 consumer health prediction tasks using prompting, fine-tuning, and context enhancement.The tasks span mental health, activity tracking, metabolism, and sleep assessment.Source: §1 · p. 2
• HealthAlpaca achieves the best performance in 8 of 10 tasks despite being substantially smaller than GPT-3.5, GPT-4, and Gemini-Pro.Source: §1 · p. 2
• Context enhancement yields up to 23.8% performance improvement, highlighting the value of contextual information in health-prediction prompts.Source: §1 · p. 2
• Case studies illustrate step-by-step reasoning by representative LLMs for time-series interpretation and personalized recommendations.Source: §1 · p. 2
• Health-LLM adapts LLMs to health prediction by prompting and training them with wearable sensor data.Source: §1 · p. 2
• The framework combines four public health datasets, ten consumer health tasks, and twelve state-of-the-art LLM evaluations.Source: §1 · p. 2

2.1. Wearable Sensor Data with LLMs

• Wearable data can be represented as raw time-series text, modality-specific or timestamp embeddings, and statistical or transformed summaries.Source: §2.1 · p. 2
• Wearable sensors provide continuous physiological time-series data that can support prediction of depression, atrial fibrillation, stress, nutrition, and stress-management outcomes.Source: §2.1 · p. 2

2.2. Health LLMs

• Health-LLM evaluates LLMs on wearable-sensor consumer health tasks, addressing a setting where domain models remain largely untested because of non-linguistic data and missing benchmarks.Source: §2.2 · p. 3
• Prior healthcare LLM work includes medical fine-tuning, mental-health adaptation from online text, and general-purpose medical evaluation, whereas this work uses multimodal wearable time-series data.Source: §2.2 · p. 3

3.1. Zero-shot Prompting

• Zero-shot prompts combine instructions, context, questions, and output formats, with user, health, and temporal context added to improve health-task grounding.Health context supplies target definitions and equations, while temporal context uses raw time-series sequences rather than aggregated statistics.Source: §3.1 · pp. 3–4

3.2. Few-shot Prompting

• Few-shot prompting uses three demonstrations and additionally applies Chain-of-Thought and Self-Consistency prompting with five candidate reasoning paths.Source: §3.2 · p. 4

3.3. Instruction Tuning

• Instruction tuning fine-tunes all parameters of a pretrained model on target health tasks so it can adapt to physiological terminology, mechanisms, and context.Source: §3.3 · p. 4
• Parameter-efficient alternatives such as LoRA train low-rank matrices while preserving computational efficiency for healthcare adaptation.Source: §3.3 · pp. 4–5

3.4. Temporal Encoding Methods

• The selected temporal encoding method is Natural Language String, chosen for interpretability, missing-value handling, and compatibility with daily, weekly, and monthly windows.The study also compares modality-specific encoding and statistical summaries.Source: §3.4 · p. 5

4.1. Datasets and Tasks

• Four wearable datasets combine physiological, behavioral, self-reported, and survey information for consumer-health prediction tasks.PMData, LifeSnaps, GLOBEM, and AW FB cover stress, readiness, fatigue, sleep, resilience, sleep disorder, anxiety, depression, activity, and calorie-related tasks.Source: §4.1 · pp. 5–6
• The benchmark includes thirteen defined consumer-health tasks classified into four topics, with optional prompt contexts used for ablation.The task table describes the task inventory and marks optional contexts for the ablation experiments.Source: Table 2 · p. 6

4.2. Models

• The study compares LLMs spanning different sizes, pre-training targets, and availability, alongside task-specific classical and pretrained-language-model baselines.Baselines include majority class, MLP, SVM, RandomForest, BERT, BioMed-RoBERTa, and PALMp-tuned models that require target-task training.Source: §4.2 · pp. 6, 8
• The cited evaluated set includes MedAlpaca, PMC-Llama, BioMedGPT, BioMistral, Asclepius, and ClinicalCamel, described as medical QA, specialized open-source, biomedical, medical-domain, clinical, and open LLMs.Source: §4.2 · pp. 6–7
• Performance is reported across ten health prediction tasks and four datasets using multiple prompting and fine-tuning techniques, with MAE, accuracy, MAPE, and F1-score metrics.The main and appendix tables also compare twelve prompted or trained LLMs with six task-specific baselines.Source: §4.2 · p. 7

5.1. Zero-shot and Few-shot Performance

• CoT-SC achieves the best performance in 9 out of 10 tasks, whereas zero-shot results show no clear model-size advantage.In zero-shot evaluation, smaller models sometimes outperform larger ones, while GPT-3.5 and GPT-4 benefit from few-shot prompting with CoT and SC in most tasks.Source: §5.1 · p. 8

5.2. Finetuning Performance

• HealthAlpaca achieves the best performance in 8 out of 10 consumer-health tasks and matches or exceeds much larger GPT-series and Gemini-Pro models.HealthAlpaca-LoRA also improves over larger models in almost all tasks, while HealthAlpaca-13B generally exceeds the 7B and parameter-efficient variants.Source: §5.2 · p. 8

5.3. Generalization of Fine-tuned Models across Datasets

• Multi-dataset fine-tuning gives HealthAlpaca reasonable cross-task generalization, while dataset-specific fine-tuning often fails on tasks from other datasets.MedAlpaca usually performs best when fine-tuned and tested on the same dataset.Source: §5.3 · pp. 8–9
• Some cross-dataset pairings, including AW FB →STRS and LifeSnaps →ANX, outperform both zero-shot and dataset-specific fine-tuning, but gains depend on overlapping dataset content.The reported synergy is not consistent across all tasks.Source: §5.3 · p. 9

5.4. Importance of different contexts in Prompt Designing for Healthcare LLMs

• Health context provides the largest average boost among the evaluated context types, while LifeSnaps reaches up to 34.46% improvement and Palmyra-Med reaches up to 44.58% with temporal context.Temporal context does not improve GPT-3.5 or GPT-4, whereas GLOBEM’s 7- and 14-day feature statistics help explain temporal-context gains.Source: §5.4 · pp. 9–10, Figure 2 · p. 10

5.5. Importance of Training Size in Fine-tuning Performance

• With 15% of the original training data, fine-tuned Alpaca already outperforms zero-shot performance on all 13 tasks, and performance increases with more data.The study evaluates 5%, 15%, 25%, 50%, and 100% training-size settings using three seeds.Source: §5.5 · p. 10

5.6.1. LLM’s capability to understand the time-series data

• HealthAlpaca uses conservative, stability-focused interpretations, whereas GPT-4 emphasizes high activity and variability when reasoning about readiness and sleep.The cases show that models can apply distinct inference strategies to the same time-series health information.Source: §5.6.1 · p. 10

5.6.2. False and Hallucinated Reasoning from LLMs

• Case studies identify false reasoning from misinterpreted data or generic standards, producing divergent readiness and sleep-disorder conclusions across models.Examples include HealthAlpaca missing nuanced sleep patterns, GPT-3.5 misreading resting-heart-rate information, and Gemini-Pro misinterpreting calorie-burn recommendations.Source: §5.6.2 · pp. 10–11

• HealthAlpaca achieved the best performance in 8 out of 10 tasks, outperforming much larger models including GPT-4 and Gemini-Pro.Source: §6 · p. 11
• Context enhancement boosted performance across all datasets and LLMs, particularly when prompts incorporated health knowledge context.Source: §6 · p. 11
• The case study exposed both reasoning capabilities and false or hallucinated reasoning in health predictions.Source: §6 · p. 11
• Reliance on self-reported health data limits clinical applicability and raises concerns about data validity and user communication.Source: §6 · p. 11
• Future work targets clinically diagnosed evaluations, ethical and regulatory compliance, explainability, and privacy-preserving technologies such as federated learning.Source: §6 · p. 11

Appendix A. Implementation Details

• Health-LLMs are fine-tuned from MedAlpaca-7b and MedAlpaca-13b models using instruction tuning with cross-entropy loss.Source: Appendix A · p. 15
• Training uses four NVIDIA A100 80GB GPUs, Adam optimization, a 2e−5 learning rate, warmup, and three or five epochs for the 7b and 13b models.Source: Appendix A · p. 15
• Prompting experiments use open-source Hugging Face models on NVIDIA A6000 GPUs and closed-source models through OpenAI and GenAI APIs.Source: Appendix A · p. 15
• Oversampling addresses class imbalance, and training sets are merged for task-specific and multi-dataset fine-tuning.Source: Appendix A · p. 15

B.0.1. Natural Language String

• A 14-day sensor sequence can be represented as daily readings of steps and resting heart rate in a language string.Source: §B.0.1 · p. 15

B.0.2. Modality-specific Encoding

• Modality-specific encoding embeds non-text modalities, including time-series data, into the same latent space as language tokens.Source: §B.0.2 · pp. 15–16
• Although modality-specific encoding may provide nuanced representations, it introduces complexity and computational overhead that may not yield performance gains.Source: §B.0.2 · p. 16

B.0.3. Statistical Summary

• Statistical summarization compresses time-series data into features such as mean, standard deviation, and median, reducing processing volume.Source: §B.0.3 · p. 16
• Statistical summaries may overlook critical temporal patterns present in raw sequences.Source: §B.0.3 · p. 16

Additional Experiments

• MAE was 0.44 ± 0.1 for stress-level prediction and 44.2 ± 7.2 for calorie-burn prediction when training and evaluation datasets were separated.These experiments used LifeSnaps labels while fine-tuning on PMData and AW_FB to assess generalizability and robustness.Source: Appendix C · p. 16

Statistical Comparisons

• Tables 6–11 report pairwise t-test results comparing baseline models with HealthAlpaca series models for Tasks 1–6: STRS, READ, FATG, SQ, SR, and SD.For Task 4, * indicates p < 0.05 and ** indicates p < 0.001.Source: Table 6 · p. 17, Table 7 · p. 17, Table 8 · p. 18

Evaluation Scope

• The appendix evaluates twelve prompted or fine-tuned LLMs and six task-specific baselines across ten health-prediction tasks and four datasets.Performance is reported with MAE, accuracy, MAPE, and F1-score, with paired t-tests comparing fine-tuned models against baselines.Source: Table 16 · p. 22, Table 17 · p. 22

Method and Case Studies

• Appendix figures document time-series encoding, zero-shot and few-shot Chain-of-Thought self-consistency exemplars, prompt decoding, and MedAlpaca fine-tuning configurations.Additional figures show case studies for readiness and sleep-disorder prediction and performance under different training configurations.Source: Figure 3 · p. 23, Figure 4 · p. 24, Figure 6 · p. 26