Skip to main content
Nubint
QUICK REVIEW

[论文解读] Patient foundation model for risk stratification in low-risk overweight patients

Zachary Flamholz, Dillon Tracy|arXiv (Cornell University)|Feb 9, 2026
Machine Learning in Healthcare被引用 0
一句话总结

患者TPP 是一种神经时序点过程模型，从诊断、检验和用药序列中学习全面、时间感知的患者表征，以预测低风险超重人群的肥胖相关健康结局和未来医疗成本。

ABSTRACT

Accurate risk stratification in patients with overweight or obesity is critical for guiding preventive care and allocating high-cost therapies such as GLP-1 receptor agonists. We present PatientTPP, a neural temporal point process (TPP) model trained on over 500,000 real-world clinical trajectories to learn patient representations from sequences of diagnoses, labs, and medications. We extend existing TPP modeling approaches to include static and numeric features and incorporate clinical knowledge for event encoding. PatientTPP representations support downstream prediction tasks, including classification of obesity-associated outcomes in low-risk individuals, even for events not explicitly modeled during training. In health economic evaluation, PatientTPP outperformed body mass index in stratifying patients by future cardiovascular-related healthcare costs, identifying higher-risk patients more efficiently. By modeling both the type and timing of clinical events, PatientTPP offers an interpretable, general-purpose foundation for patient risk modeling with direct applications to obesity-related care and cost targeting.

研究动机与目标

  • 开发一个患者表示框架（PatientTPP），通过神经时序点过程对不规则的临床轨迹进行建模。
  • 在 TPP 中加入静态协变量和数值协变量，以及领域知情事件编码，以获得更丰富的表征。
  • 评估 PatientTPP 表征向肥胖相关结局和健康经济学指标的迁移能力。
  • 证明基于 PatientTPP 的风险分层可以在 BMI 基础方法之外更好地识别高成本患者。

提出的方法

  • 在第一步中将静态协变量（性别、种族、伪年龄）注入，以扩展基于 AttNHP 的神经 TPP。
  • 为数值特征（如 BMI、HbA1c）添加并行的 TPP 流，离散化成离散桶。
  • 引入预训练临床嵌入以丰富事件表征。
  • 在包含 426,039 条患者轨迹、64+ 个事件的数据上训练，优化事件的对数似然和事件间时序的生存项。
  • 通过从预测事件概率派生固定长度的患者嵌入，用于下游任务实现迁移学习。
  • 使用下游逻辑回归预测肥胖相关结局、线性回归预测成本，并从 PatientTPP 嵌入中进行迁移学习来评估。
Figure 1: A. A temporal point process is an event sequence defined over some interval $[0,T)$ . Events arrive at real-valued timestamps on this interval. Events may have classes associated with them (also known as “marks”) but not magnitudes (at least not out-of-the-box). Events have no duration. Th
Figure 1: A. A temporal point process is an event sequence defined over some interval $[0,T)$ . Events arrive at real-valued timestamps on this interval. Events may have classes associated with them (also known as “marks”) but not magnitudes (at least not out-of-the-box). Events have no duration. Th

实验结果

研究问题

  • RQ1一个神经时序点过程是否能够从包含静态和数值特征的不规则临床轨迹中学习出鲁棒且可迁移的患者表征？
  • RQ2PatientTPP 嵌入是否在低风险超重人群中提升对肥胖相关健康结局和医疗成本的预测，超越传统指标如 BMI？
  • RQ3学习到的表征是否可迁移到培训阶段未显式建模的未见结局？
  • RQ4PatientTPP 表征是否支持以成本为导向的高成本干预风险分层（如 GLP-1 疗法）？

主要发现

ConditionPatientTPPWeighted guessing
None5363543462
Sleep Apnea3029147
Arrhythmias excl. atrial fibrillation182754
Primary CKD stage 1–4141020
COPD134218
Atrial fibrillation95214
Acute Kidney Failure6828
Cholelithiasis5404
Patient death5309
Ischemic stroke4964
Acute Myocardial Infarction4196
Cardiomyopathy3971
Cholecystitis1580
Cardiac Arrest1490
Breast cancer1050
Acute Pancreatitis990
Acute Ischemic Heart Disease excl. MI980
Hemorrhagic stroke720
Colorectal cancer420
Kidney cancer280
CKD 5 and ESRD250
Pancreatic cancer250
Endometrial cancer240
Thyroid cancer220
Hepatocellular carcinoma190
Multiple myeloma130
Ovarian cancer90
Complication of MI80
Subsequent MI80
Stomach cancer80
Esophageal cancer63
Gallbladder cancern < 50
  • 与基于 BMI 的分层相比，PatientTPP 在肥胖相关结局和成本相关预测方面的 AUROC 有所提升（成本分析中显示 CV-cost AUROC 提升）。
  • PatientTPP 嵌入使得在训练阶段未显式建模的若干与肥胖相关的结局也能被预测，迁移学习下的 AUROC 对不同结局有所波动。
  • 在成本分层中，基于 PatientTPP 的 CV-cost 排名覆盖了前 25% 患者的总成本的 39.3%，而基于 BMI 的排名为 22.7%（p<1e-5）。
  • 一个 6-队列的迁移学习设置显示各结局的 AUROC 在 0.53 到 0.79 之间，表明学习表征中存在预测信号。
  • PatientTPP 表征能够识别潜在需要 GLP-1 疗法的负担较高的队列，在所考察队列中四年内成本差异最高可达约 8090 万美元。
Figure 2: A. PatientTPP extensions to the AttNHP encoder architecture. Input sequence events and timestamps are embedded separately, then concatenated and fed to a multi-head self-attentive transformer. Invariant features are added in the main context-building loop are are nonzero for the first item
Figure 2: A. PatientTPP extensions to the AttNHP encoder architecture. Input sequence events and timestamps are embedded separately, then concatenated and fed to a multi-head self-attentive transformer. Invariant features are added in the main context-building loop are are nonzero for the first item

更好的研究，从现在开始

从论文设计到论文写作，大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成，并经人工编辑审核。