NCI Data Jamboree (Project Abstract Submission): Submission #35

Submission information
Submission Number: 35
Submission ID: 188978
Submission UUID: 114f69d4-b322-46e2-a462-8577ea09e926

Created: Sat, 07/25/2026 - 08:36
Completed: Sat, 07/25/2026 - 08:41
Changed: Sat, 07/25/2026 - 08:41

Remote IP address: 10.208.24.192
Submitted by: Anonymous
Language: English

Is draft: No
First Name Xiaozhong
Middle Initial
Last Name Liu
Degree(s) Ph.D.
Position/Title/Career Status Professor
Organization Worcester Polytechnic Institute
Organization Address Worcester
Email xliu14@wpi.edu
List of Additional Authors
  • First Name: Patricia
    Last Name: Mabry
    Affiliation: HealthPartners
  • First Name: Dylan
    Last Name: Zylla
    Affiliation: HealthPartners
  • First Name: Jinhong
    Last Name: Yu
    Affiliation: Worcester Polytechnic Institute
  • First Name: Zhikai
    Last Name: Xue
    Affiliation: Worcester Polytechnic Institute
Abstract Category Employing statistical, computational, and informatics tools, algorithms, and methods to integrate or analyze data
Abstract Keywords Immune checkpoint inhibitor–associated myocarditis; Cohort-aware AI; Adverse event surveillance; Longitudinal patient monitoring; Clinical informatics
Abstract Title Onco-Guard: Cohort-Aware AI for Earlier Detection of Immune Checkpoint Inhibitor–Associated Myocarditis
Abstract Patients receiving immune checkpoint inhibitors spend most of their treatment time outside the clinic, while rare but rapidly progressive toxicities may emerge between scheduled visits. Immune checkpoint inhibitor-associated myocarditis is a compelling proof-of-concept because early symptoms and biomarker changes can be nonspecific, fragmented, and inconsistently sampled. We propose Onco-Guard, a cohort-aware AI system that evaluates each patient against both the patient’s own longitudinal baseline (self-as-reference) and clinically similar patients receiving comparable regimens (cohort-as-reference). The system will organize clinical, molecular, laboratory, symptom, and physiologic information into reviewable longitudinal trajectories and produce an explicit evidence trace, missing-data indicators, and a prioritized list of patients warranting closer clinical review; it will not make autonomous clinical decisions.

For the Jamboree, we will investigate whether available NCI and dbGaP resources can support rigorous landmark-based evaluation of earlier toxicity signals. Candidate datasets include phs003413 (checkpoint myocarditis), phs003284 (immune-related adverse events after checkpoint blockade in sarcoma), and phs003412 (Lung-MAP S1400I). Before and during the Jamboree, we will assess variable availability, harmonize usable timelines, define myocarditis phenotypes and recognition timepoints with clinical experts, and compare threshold-based, longitudinal, representation-learning, and agent-assisted approaches while masking all post-landmark information. Evaluation will emphasize event counts, lead time, false-positive burden, uncertainty, and limitations caused by rarity and missing data.

Our team brings expertise in multimodal AI, patient memory, wearable monitoring, clinical translation, and oncology toxicity workflows, supported by an existing Onco-Guard prototype running on simulated cohorts. Clinical collaborators at HealthPartners will guide phenotype definition, workflow relevance, and human-in-the-loop review. We seek cardio-oncology expertise for phenotype adjudication and NCI data-commons expertise for timeline harmonization. The Jamboree deliverables will be a reusable longitudinal schema, an honest data-readiness and gap analysis, a landmark evaluation pipeline, and a clinically reviewable demonstration that will inform a planned ITCR U01 application.