Large language models in medical time series analysis

Yu Han; Cigdem Beyan; Xiang Zhang; Xiaofeng Liu; Nan Liu; Jimeng Sun; Shenda Hong; Cheng Ding

eess.SPSep 7, 2026

Large language models in medical time series analysis

Authors: Yu Han, Cigdem Beyan, Xiang Zhang, Xiaofeng Liu, Nan Liu, Jimeng Sun, Shenda Hong, Cheng Ding,

Organizations: AI for Good (AIGO), Istituto Italiano di Tecnologia, Genoa, Liguria, Italy · DITEN, University of Genoa, Genoa, Liguria, Italy · Department of Computer Science, University of Verona, Verona, Veneto, Italy · Department of Computer Science, University of North Carolina at Charlotte, Charlotte, North Carolina, USA · Department of Biomedical Informatics and Data Science, Yale University, New Haven, Connecticut, USA · NUS Artificial Intelligence Institute, National University of Singapore, Singapore, Singapore · Siebel School of Computing and Data Science, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA · Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA · Keiji AI, Seattle, Washington, USA · National Institute of Health Data Science, Peking University, Beijing, China · Institute for Artificial Intelligence, Peking University, Beijing, China · Nanjing University of Aeronautics and Astronautics, College of Artificial Intelligence, Nanjing, Jiangsu, China

Abstract

Medical time series (MedTS), including electrocardiograms (ECG), electroencephalograms (EEG), photoplethysmography (PPG), and vital-sign recordings, are central to clinical diagnosis and health monitoring. As large language models (LLMs) have advanced, a growing body of work has examined how their reasoning, generation, and knowledge-integration capabilities can support MedTS analysis. Yet existing studies remain scattered, and the field still lacks a clear view of how these models should be designed, integrated into clinical workflows, and evaluated. This review synthesizes recent work on large language models for medical time series analysis (MedTSLLMs), covering both methodological progress and issues related to real-world deployment. We review model architectures, data resources, and processing pipelines, and prompt design strategies adapted for diverse clinical scenarios. We further organize existing MedTS applications, ranging from diagnostic interpretation and report generation to longitudinal health monitoring and physiological signal synthesis, highlighting task-specific design choices, common evaluation protocols, and empirical findings reported across studies. By bringing together current practices and open challenges, this review aims to provide a clearer foundation for developing, evaluating, and deploying MedTSLLMs responsibly in healthcare. We also maintain a regularly updated list of MedTSLLM studies and resources at: https://github.com/hy727/MedTSLLM-Review.

Figures & tables

Review	Focus	Multimodal Integration	Architecture	Pipeline	LLM Role	Prompt Design	Evaluation Protocol	Clinical Guidance	Task Scope
Zhou et al. [ 266 ]	Medicine	$\checkmark$	$\checkmark$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\checkmark$	D+G+H
Ansari et al. [ 8 ]	ECG	$\checkmark$	$\checkmark$	$\times$	$\times$	$\times$	$\checkmark$	$\times$	D+G
Khan et al. [ 98 ]	ECG	$\checkmark$	$\checkmark$	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$	D+G
Babu et al. [ 12 ]	EEG	$\checkmark$	$\checkmark$	$\times$	$\checkmark$	$\times$	$\times$	$\times$	D+G+S
Arava et al. [ 10 ]	Wearable	$\checkmark$	$\times$	$\checkmark$	$\checkmark$	$\times$	$\checkmark$	$\checkmark$	G+H
Ding et al. [ 46 ]	MedTS	$\times$	$\checkmark$	$\times$	$\times$	$\times$	$\times$	$\times$	D+G+H+S

Table 1: Comparison of our review with existing review papers in LLM for biosignals. Abbreviation: MedTS: medical time series. D: diagnosis. G: generation. H: health monitoring. S: medical time-series synthesis.

Figure 1: Overview of MedTSLLM workflow. Here we outline the workflow of MedTSLLMs from multi-source MedTS acquisition and signal streams to LLM-based pipeline, knowledge integration, and downstream clinical applications.

Model Structures	Models	# Params	Pre-training Data Scale	Representative Applications	Key Strengths & Limitations
Encoder-only	BERT [ 44 ]	110M/340M	3.3B tokens	Arrhythmia diagnosis; Sleep apnea detection; Sentiment analysis; Reading comprehension	Strengths: Bidirectional contextual modeling and efficient representation learning. Limitations: Limited generation and instruction-following capabilities; task-specific heads are often required.
	ClinicalBERT [ 169 ]	$\sim$ 110M	880M words
	BiolinkBERT [ 246 ]	$\sim$ 110M/333M	21GB text
	BEiT [ 14 ]	86M/304M	14M images
	Ada-v2 [ 162 ]	-	-
	ViLT [ 103 ]	$\sim$ 87M	5.1M image-text pairs

Table 2: Summary of the LLMs used in existing MedTS studies, including their underlying structures, numbers of parameters, pre-training data scales, representative applications, and key strengths and limitations. Column “# Params” denotes the number of parameters. Abbreviations: K: thousand; M: million; B: billion; T: trillion; GB: gigabyte; IFT: instruction fine-tuning data.

Paradigm	Parameter size	Computational Cost	Data Requirements	Key Strengths & Limitations	Representative Applications
SFT	Large	High	Task-specific labeled signal-text pairs	Strengths: Strong task specialization. Limitations: Data-intensive and prone to overfitting on limited datasets.	Seizure classification; Arrhythmia diagnosis; ECG interpretation
IFT	Large	High	High-quality instruction- response pairs	Strengths: Strong instruction following and controllable outputs. Limitations: Sensitive to instruction quality and coverage.	Medical QA; Health assessment; Clinical report generation
PEFT	Small	Low	SFT/IFT data	Strengths: Compute- and memory-efficient. Limitations: Limited adaptation capacity and configuration-sensitive.	Medical QA; EEG-to-text translation; Clinical report generation

Table 3: Comparison of common fine-tuning paradigms for MedTSLLMs in terms of trainable parameter size, computational cost, data requirements, key strengths and limitations, and representative applications.

Figure 2: Diagram of the prompt templates used in MedTSLLM studies. Here we take ECG-based arrhythmia diagnosis as an example to show how to design a prompt template under different prompt settings. Due to the space limitation, we use some abbreviations in this figure. Abbreviations: MedTS: medical time-series. TS: time-series. AFib: atrial fibrillation. PAC: premature atrial contractions. PVC: premature ventricular contractions. HC: healthy control. User: user profile. Q: question. DK: domain knowledge. To help readers assess the maturity and current adoption of different prompt designs, we report the number of representative studies using each template in the surveyed literature: Type 1: Basic Prompt Template (

n=10

), Type 2: User-Profile Augmented Template (

n=9

), Type 3: Domain Knowledge-Enhanced Template (

n=11

), Type 4: Retrieval-Augmented Generation Template (

n=3

), Type 5: Label-Injected Template (

n=5

), and Type 6: Soft Embedding Learnable Template (

n=2

Figure 3: Integrated overview of potential applications of MedTSLLMs. Here, we briefly present six main applications of MedTSLLMs and outline their frameworks and key components of MedTSLLMs in different clinical scenarios. Abbreviations: AF: atrial fibrillation. STEMI: ST-segment elevation myocardial infarction. HCM: hypertrophic cardiomyopathy. HTN: hypertension. BP: blood pressure. HR: heart rate.

Application	Method	Architecture	Role of LLM	# Params	Prompt Technique	Prompt Template	Dataset	Data Scale	Data Type	Evaluation (Task: Score)
Medical disease diagnosis	1 1 footnotemark: 1 HeartBEiT [ 210 ]	BEiT	Backbone	86M	N/A	N/A	MSHS PTB-XL	8.5M samples 2.1M patients	ECG	LVEF diagnosis: 0.86 AUROC HCM diagnosis: 0.77 AUROC STEMI diagnosis: 0.88 AUROC
	sEHR-ECG-Text [ 107 ]	BERT	Text encoder	15M	N/A	N/A	Mayo EHRs PhysioNet2020 Chapman	9M samples 2.4M patients	ECG EHRs sEHRs	Arrhythmia diagnosis: 0.933 AUROC CAS diagnosis: 0.892 AUROC Myocarditis diagnosis: 0.905 AUROC CAM diagnosis: 0.961 AUROC PH diagnosis: 0.917 AUROC LVEF diagnosis: 0.933 AUROC
	ECGBERT [ 34 ]	BERT	Backbone	N/A	N/A	N/A	MIMIC-III PTB-XL CPSC2018 Georgia MIT-BIH Apnea-ECG	88K samples 29K patients	ECG	Arrhythmia diagnosis: 0.973 ACC Sleep apnea detection: 0.725 ACC
	METS [ 119 ]	ClinicalBERT	Text encoder	N/A	N/A	N/A	MIMIC-III PTB-XL MIT-BIH	22K samples	ECG Clinical reports	Abnormality diagnosis: 0.842 ACC Arrhythmia diagnosis: 0.746 ACC
	2 2 footnotemark: 2 MERL-CKEPE [ 124 ]	GPT-4	Prompt generator	1.2B	Clinical-guided	Ins+Q+GT $\rightarrow$ Prompts	MIMIC-ECG PTB-XL CPSC2018 CSN	770K pairs	ECG Clinical notes	Abnormality diagnosis: 0.906 AUC Zero-shot ECG classification: 0.829 AUC
	Zero-shot RAG [ 250 ]	LLaMA-2 GPT-3.5	Text decoder	7B 13B 175B	Clinical-guided	Ins+TN+DK $\rightarrow$ Labels+Interp	PTB-XL+ Apnea-ECG	56K samples	ECG	Arrhythmia diagnosis: 0.669 F1 Sleep apnea detection: 0.801 F1

Table 4: Summary of existing MedTSLLMs tailored to various clinical applications, in terms of their architecture, functional role, the number of parameters, the prompt techniques and templates, the source, scale, and detailed type of the pre-training/fine-tuning data, and the evaluation information, including the task and performance. Note that, in this review, arrhythmia diagnosis refers to rhythm-related disorders, whereas abnormality diagnosis denotes non-arrhythmic general abnormalities, such as myocardial infarction and ST/T changes. Specific cardiac diseases or functional conditions, such as LVEF and STEMI, are listed separately in the Evaluation column when applicable. Abbreviations: M: million. B: billion. TS: time-series. QA: question-answering. CoT: chain-of-thought prompting. ToT: tree-of-thought prompting. Ins: instruction. Q: question. GT: ground truth. TN: textualized numeric context. DK: domain knowledge. Interp: Interpretation. User: user profile. CEN: clinical encounter notes. DTN: descriptive textual narratives. ImgEmb: image token embedding. TSEmb: time-series token embedding. TEXmb: text token embedding. ObjEmb: object-label token embedding. SE: soft embeddings. RR: retrieved reports. DxL: diagnosis labels. TF: time-frequency. BEL: brain electrode location. TW: target words. SHGC: structured hypergraph context. BP: blood pressure. HR: heart rate. CAS: Coronary atherosclerosis. CAM: Cardiac amyloidosis. PH: pulmonary hypertension. LVEF: left ventricular ejection fraction. HCM: hypertrophic cardiomyopathy. STEMI: ST-segment elevation myocardial infarction. AF: atrial fibrillation. IBI: interbeat interval. CBP: cuffless blood pressure

Figure 4: Categories of MedTSLLM applications in the reviewed papers.

Medical LLMs

Medical large language models are AI systems adapted to clinical and biomedical data for health reasoning, information extraction, and communication.

See all 397 papers

HEARTS: Benchmarking LLM Reasoning on Health Time Series
Feb 25, 2026Sirui Li, Shuhan Xiao, Mihir Joshi +4LLM Evaluation Time Series Reasoning
University of California, Los Angeles, 2Google Research · 2Google Research
Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis
Aug 19, 2026Yihan Xie, Hanwen Cui, Runze Ye +8Electrocardiography Electrocardiogram Interpretation
Zhejiang University · Beijing Institute of Technology · University of Electronic Science and Technology of China
MedTS-TTT: Test-Time Training for Medical Time Series Classification
Jun 19, 2026Mingzhi Chen, Yiyu Gui, Guibo LuoTime Series Classification Electrocardiogram Classification
Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University, Shenzhen, China

Feb 25, 2026J/K move · Enter open · S save

HEARTS: Benchmarking LLM Reasoning on Health Time Series

Sirui Li, Shuhan Xiao, Mihir Joshi, Ahmed Metwally+3

University of California, Los Angeles, 2Google Research · 2Google Research

LLM Evaluation Time Series Reasoning Healthcare

Open paper AskPDF arXiv

The rise of large language models (LLMs) has shifted time series analysis from narrow analytics to general-purpose reasoning. Yet, existing benchmarks cover only a small set of health time series modalities and tasks, failing to reflect the diverse domains and extensive temporal dependencies inherent in real-world physiological modeling. To bridge these gaps, we introduce HEARTS (Health Reasoning over Time Series), a unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series. HEARTS integrates 16 real-world datasets across 12 health domains and 20 signal modalities, and defines a comprehensive taxonomy of 110 tasks grouped into four core capabilities: Perception, Inference, Generation, and Deduction. Evaluating 16 state-of-the-art LLMs on more than 20K test samples reveals intriguing findings. First, LLMs substantially underperform specialized models, and their performance is only weakly related to general reasoning scores. Moreover, LLMs often rely on simple heuristics and struggle with multi-step temporal reasoning. Finally, performance declines with increasing temporal complexity, with similar failure modes within model families, indicating that scaling alone is insufficient. By making these gaps measurable, HEARTS provides a standardized testbed and living benchmark for developing next-generation LLM agents capable of reasoning over diverse health signals.

Abstract

Medical LLMs

Explore similar work

HEARTS: Benchmarking LLM Reasoning on Health Time Series

Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis

MedTS-TTT: Test-Time Training for Medical Time Series Classification