Clinical Validation of a Pre-trained and Fine-tuned Model for Analyzing Time-Series Electronic Health Records in Critical Care

The goal of this observational study is to validate the generalization capability and predictive performance of the "Time-Series Electronic Health Record Foundation Model (EHR Foundation Model)", which was pre-trained on a large-scale critical care dataset, within the real-world clinical environment of the Intensive Care Unit (ICU) at Peking University People's Hospital. The main question it aims to answer is:

· Based on the "pre-training + fine-tuning" paradigm, whether the EHR foundation model can effectively provide accurate and dynamic prognostic support information in real-world clinical scenarios.

Participants who agree to take part in the research, in addition to completing the signed informed consent form, if the patient's hospital stay is too short, will be obtained information on prognostic outcomes (e.g., 28-day mortality) through telephone follow-up.

Study Overview

• Status: Recruiting
• Conditions: Mortality Prediction; Critical Care Medicine

Detailed Description

The clinical validation design combines a retrospective cohort study with a prospective cohort study.

• Study Type: Observational
• Enrollment (Estimated): 4800

Contacts and Locations

• Study Contact: Name: Huiying Zhao, PhD; Phone Number: +86 13811088270; Email: zhaohuiying@pku.edu.cn
• Study Contact Backup: Name: Xiaojiang Liu, M.Med; Phone Number: +86 13717952776; Email: liuxiaojiang@pku.org.cn

Study Locations

• China
  • Beijing Municipality
    • Beijing, Beijing Municipality, China, 100044
      • Recruiting
      • Peking University People's Hospital
      • Contact: Huiying Zhao, PhD; Phone Number: +86 13811088270; Email: zhaohuiying@pku.edu.cn
      • Principal Investigator: Huiying Zhao, PhD
      • Sub-Investigator: Xiaojiang Liu, M.Med
      • Sub-Investigator: Chenxiao Hao, M.Med

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: No

• Sampling Method: Non-Probability Sample

Study Population

Study Setting: Department of Critical Care Medicine (ICU), Peking University People's Hospital.

Data Source and Time Period:

Retrospective Validation Cohort: Electronic health record data from all patients admitted to the ICU of Peking University People's Hospital between January 1, 2009 and December 31, 2025. This dataset will be used for local fine-tuning and internal validation of the model.

Prospective Validation Cohort: Newly diagnosed cases admitted to the ICU of Peking University People's Hospital from March 1, 2026 to December 31, 2026 (following project initiation). This dataset will be used for external independent validation of the model.

Description

Inclusion Criteria:

• Age ≥ 18 years, no gender restriction;
• ICU length of stay ≥ 24 hours (to ensure sufficient time-series monitoring data for model generation);
• Complete electronic health record data, including baseline demographic characteristics and at least one complete laboratory test record after ICU admission.

Exclusion Criteria:

• Patients who are transferred out or die within 24 hours of ICU admission;
• Duplicate admission records for non-initial ICU admissions (only the initial ICU admission record is retained to ensure independence);
• Severe deficiency in core data (e.g., absence of major vital sign recordings or > 50% missing key laboratory test results);
• Patients with abandonment of treatment or discharge against medical advice, leading to inability to ascertain the definitive clinical outcome.

Study Plan

How is the study designed?

Number of groups / cohorts

2

Cohorts and Interventions

Group / Cohort
Non-survivor group; Patients confirmed to have died within 28 days (outcome obtained from hospital records or telephone follow-up).
Survivor group; Patients confirmed to have survived beyond 28 days (outcome obtained from hospital records or telephone follow-up).

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
In-hospital Mortality	Death events occurring during the patient's current hospitalization.	The endpoints are patient discharge or death. For prognostic outcomes (e.g., 28-day mortality), if the patient's hospital stay is insufficient, the outcome will be obtained through telephone follow-up.

Collaborators and Investigators

• Sponsor: Peking University People's Hospital
• Collaborators: Peking University

Publications and helpful links

General Publications

• Mao Q, Jay M, Hoffman JL, Calvert J, Barton C, Shimabukuro D, Shieh L, Chettipally U, Fletcher G, Kerem Y, Zhou Y, Das R. Multicentre validation of a sepsis prediction algorithm using only vital sign data in the emergency department, general ward and ICU. BMJ Open. 2018 Jan 26;8(1):e017833. doi: 10.1136/bmjopen-2017-017833.
• MUHAMMAD G, ALSHEHRI F, KARRAY F, et al. A comprehensive survey on multimodal medical signals fusion for smart healthcare systems [J]. Information Fusion, 2021.
• Amugongo LM, Mascheroni P, Brooks S, Doering S, Seidel J. Retrieval augmented generation for large language models in healthcare: A systematic review. PLOS Digit Health. 2025 Jun 11;4(6):e0000877. doi: 10.1371/journal.pdig.0000877. eCollection 2025 Jun.
• MA L, GAO J, WANG Y, et al. Adacare: Explainable clinical health status representation learning via scale-adaptive feature extraction and recalibration; proceedings of the Proceedings of the AAAI Conference on Artificial Intelligence, F 2020].
• MA M, REN J, ZHAO L, et al. Smil: Multimodal learning with severely missing modality; proceedings of the Proceedings of the AAAI conference on artificial intelligence, F, 2021 [C].
• Chen J, Qi TD, Vu J, Wen Y. A deep learning approach for inpatient length of stay and mortality prediction. J Biomed Inform. 2023 Nov;147:104526. doi: 10.1016/j.jbi.2023.104526. Epub 2023 Oct 17.
• SURESH H, GONG J J, GUTTAG J V. Learning tasks for multitask learning: Heterogenous patient populations in the icu; proceedings of the Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, F 2018].
• Meyer ML, Fitzgerald BG, Paz-Ares L, Cappuzzo F, Janne PA, Peters S, Hirsch FR. New promises and challenges in the treatment of advanced non-small-cell lung cancer. Lancet. 2024 Aug 24;404(10454):803-822. doi: 10.1016/S0140-6736(24)01029-8. Epub 2024 Aug 6.
• Jackson VA, Emanuel L. Navigating and Communicating about Serious Illness and End of Life. N Engl J Med. 2024 Jan 4;390(1):63-69. doi: 10.1056/NEJMcp2304436. Epub 2023 Dec 20. No abstract available.
• MA L, ZHANG C, WANG Y, et al. Concare: Personalized clinical feature embedding via capturing the healthcare context; proceedings of the Proceedings of the AAAI Conference on Artificial Intelligence, F 2020].

Study record dates

Study Major Dates

• Study Start (Actual): January 1, 2026
• Primary Completion (Estimated): December 31, 2026
• Study Completion (Estimated): December 31, 2027

Study Registration Dates

• First Submitted: August 26, 2026
• First Submitted That Met QC Criteria: August 26, 2026
• First Posted (Actual): August 31, 2026

Study Record Updates

• Last Update Posted (Actual): August 31, 2026
• Last Update Submitted That Met QC Criteria: August 26, 2026
• Last Verified: March 1, 2026

More Information

Terms related to this study

• Keywords: Intensive Care Units; Validation Study; Electronic Health Records; Hospital Mortality; Models, Theoretical
• Other Study ID Numbers: 2026PHB018-001

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No