LifeLong Digital Twin: A Unified Modeling Paradigm and Agent Harness for Event-Driven Lifelong Health State Trajectories

Jin Jiang; Sean Yates; Jasper Chong; Raymond Brooks; Alex Lawson; Yuqin Qin; Liangcai Gao

cs.AIOct 4, 2026

LifeLong Digital Twin: A Unified Modeling Paradigm and Agent Harness for Event-Driven Lifelong Health State Trajectories

Authors: Jin Jiang, Sean Yates, Jasper Chong, Raymond Brooks, Alex Lawson, Yuqin Qin, Liangcai Gao

Organizations: Peking University · VITA-Med Group

Abstract

Human health is a continuous, dynamic trajectory shaped by the cumulative interplay of biological processes, clinical events, behaviors and environmental exposures across the life course. Unifying the full breadth of lifelong health information, including longitudinal records, genetic variation, molecular profiles and environmental histories, is essential for whole-person modeling and remains a major challenge. We introduce LifeLong Digital Twin, a unified, event-driven modeling paradigm that organizes Life Events into daily Health States and accumulates them into Lifelong Health Context. An accompanying Agent Harness incorporates multimodal evidence beyond the language model's textual context. We evaluate four language models across 25 disease endpoints on three tasks: Disease Trajectory Forecasting, Disease Risk Ranking and Multi-horizon Disease Prediction. The approach yields marked gains over the reference condition: model-averaged F1 increases by 22.0% for disease identification in trajectory forecasting and 18.3% for five-year disease outcomes; thyroid-disease F1 reaches 0.669. The framework provides a foundation for whole-person digital twins and research on personalized lifelong disease prevention.

Figures & tables

Figure 1: From Life Events to Lifelong Health Context. a , Seven synthetic Life Events are grouped into three daily Health States containing two, two, and three events. The ordered states form an accumulated history, expressed as chronological event text for an LLM. Earlier events remain in the history without being copied into later states; ellipses indicate continuation. b , Six event categories and representative finer types. Colors link events to categories; the equal-sized sectors organize the taxonomy and do not indicate data proportions. The same illustrative history is expanded in Figure 2 a.

Figure 2: Direct reasoning with Lifelong Health Context and the Agent Harness. a , Chronological event text preserves dated Health States and is supplied with a task to an LLM. b , The Agent Harness retains the same complete textual Context while connecting a file-based Health State Workspace, a runtime hosting the Health Reasoning Agent, and Specialized Skills. Accepted evidence can enrich subsequent Context; predictions remain separate task outputs. The example is synthetic. File cards illustrate logical organization rather than a prescribed file per day; Skills show representative capabilities.

		A: trajectory forecasting		B: ranking	C: prediction
Model	Condition	Disease F1	Within-1-year F1	NDCG@5	Micro-F1
GPT-5.5	Baseline	0.4433	0.3548	0.6222	0.4826
	Context (ours)	0.5121	0.4416	0.6836	0.5566
	Harness (ours)	0.5249	0.4504	0.6903	0.5644
GPT-5.6 Terra	Baseline	0.3737	0.3184	0.5600	0.4669
	Context (ours)	0.4605	0.4013	0.6249	0.5275

Table 1: Five-year primary scores. Context and Harness are two formulations of our framework.

Figure 3: Overall performance of Context and Harness relative to Baseline. a–c, Reported five-year primary scores for each model; task-specific focused axes retain absolute values. d, Score-point gains over Baseline for identification, year agreement, ranking and cumulative prediction. A uses exact endpoints; B/C use families. Values are point estimates; Ctx denotes Context and H denotes Harness.

Figure 4: Performance across six cumulative horizons. a–b, Three-condition curves for every model, using family NDCG@5 for B and family micro-F1 for C. Horizons are shown as equally spaced categories; y-ranges are shared within each task. c, Gains over each model’s Baseline, with a common symmetric ±15-point color scale. Negative values are retained. B uses horizon-specific scored populations; curves do not track a fixed cohort.

Figure 5: Disease coverage and temporal variation. All values are equal-weight means of four model-specific scores, not pooled participant estimates. a, Context and Harness F1 gains over Baseline for all 25 exact endpoints at five years, with absolute scores alongside. Improvement counts use unrounded values. b, All 21 C families across six horizons: circle area is absolute F1, and color is the gain over Baseline on a shared ±80-point scale; crosses mark zero F1. The five cancer endpoints overlap in a and form one family in b. Disease supports and paired uncertainty are unavailable.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

ID	Endpoint	ICD-10 definition
Circulatory disorders
D01	Essential (primary) hypertension	I10
D02	Chronic ischaemic heart disease	I25, excluding I25.2
D03	Atrial fibrillation and flutter	I48
D04	Heart failure	I50
D05	Pulmonary hypertension / chronic cor pulmonale	I27

Appendix

Table 2: Disease endpoint dictionary: nine categories and 25 endpoints. Ranges include their component codes unless an exclusion is stated.

Explore similar work

Marrying Generative Model of Healthcare Events with Digital Twin of Social Determinants of Health for Disease Reasoning
May 10, 2026Ziquan Wei, Tingting Dan, Guorong WuReadmission Prediction Healthcare
Department of Computer Science 2Department of Psychiatry University of North Carolina at Chapel Hill
LongAgent: History-Guided Agentic Search for Longitudinal Outcome Prediction
Sep 14, 2026Siyao Wang, Florian Guitton, Shuojie Fu +3Long-Horizon Agents Longitudinal
Data Science Institute, Imperial College London, London, UK · Department of Computing, Imperial College London, London, UK · Department of Brain Sciences, Imperial College London, London, UK
Simulating clinical interventions with a generative multimodal model of human physiology
Apr 30, 2026Guy Lutsker, Gal Sapir, Jordi Merino +7Patient Trajectories Physiological Data
Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel. · Department of Molecular Cell Biology, Weizmann Institute of Science, Rehovot, Israel. · NVIDIA, Tel Aviv, Israel. +5

May 10, 2026, cs.AIJ/K move · Enter open · S save

Marrying Generative Model of Healthcare Events with Digital Twin of Social Determinants of Health for Disease Reasoning

Ziquan Wei, Tingting Dan, Guorong Wu

Department of Computer Science 2Department of Psychiatry University of North Carolina at Chapel Hill

Readmission Prediction Healthcare International Classification Of Disease Coding Generative Models Digital Twins Latent Diffusion Model

Open paper AskPDF arXiv

Despite the central role of sensor-derived measurements such as imaging traits and plasma biomarkers in biomedical research and clinical practice, existing generative models for disease prediction largely depend on event-level representations from hospital and registry data. Given the multi-factorial nature of human disease, the absence of explicit modeling of social determinants of health (SDoH), even in the limited form of ICD-coded proxies (chapters Z and V--Y in ICD-10), limits the capacity for personalized disease modeling and clinical decision support. To address this limitation, we propose a generative model with ICD-coded proxies of SDoH for \textit{in silico} modeling of disease reasoning, a conditioned latent diffusion framework that establishes the connection between multi-organ sensor data with tokenized healthcare events. Specifically, we introduce a novel geometric diffusion model to characterize the temporal evolution of complex data representation such as brain networks (region-to-region connectivity encoded in a graph), in parallel with diffusion models for tabular data from other organ systems. Together, we integrate the generative model with digitalized SDoH proxies (coined \modelname{}) for simulated intervention and reasoning of future disease trajectories. We conduct extensive experiments on the UK Biobank (UKB) dataset, which contains organ-specific imaging traits, including brain (44,834), heart (23,987), liver (28,722), and kidney (32,155), along with nearly 500k medical history sequences (age range: 25

\sim

89 years). Our \modelname{} achieves significant improvements over state-of-the-art human disease autoregressive models and imaging trait generative baselines.

Abstract

Explore similar work

Marrying Generative Model of Healthcare Events with Digital Twin of Social Determinants of Health for Disease Reasoning

LongAgent: History-Guided Agentic Search for Longitudinal Outcome Prediction

Simulating clinical interventions with a generative multimodal model of human physiology