NCI Data Jamboree (Project Abstract Submission): Submission #49

Presenter Information
---------------------
First Name: Nikita
Middle Initial: {Empty}
Last Name: Thakur
Degree(s): Master of Science, Computer Science
Position/Title/Career Status: Research Data Analyst (Analytics and AI lead)
Organization: University of Illinois Cancer Center
Organization Address:
Chicago

Email: nthaku3@uic.edu

Additional Authors
------------------
List of Additional Authors:
- First Name: Saransh
  Last Name: Singh
  Affiliation: University of Illinois Cancer Center
- First Name: Juhi
  Last Name: Anand
  Affiliation: University of Illinois Cancer Center
- First Name: Lakshmi Sravya
  Last Name: Rachakonda
  Affiliation: University of Illinois Cancer Center


Abstract Information
--------------------
Abstract Category: Evaluating data quality for reproducibility and AI-readiness
Abstract Keywords:  AI-readiness; data quality; Cancer Research Data Commons; multimodal oncology data; data leakage
Abstract Title: AI-Readiness Scorecard: A Task-Relative Profiler for Multimodal Cancer Research Data Commons Datasets
Abstract:
Rationale: NIH Bridge2AI criteria establish that FAIR conformance alone does not make a dataset AI-ready, yet the criterion most likely to invalidate a downstream model, Characterization, remains largely unautomated. Researchers assembling oncology AI cohorts evaluate batch effects, informative missingness, leakage risk, and clinical-variable completeness ad hoc and per repository, because CRDC Commons were developed independently with distinct data models.
Objective: Prototype an installable Python profiler answering one question: Does this oncology dataset support this AI task, and if not, what is missing?
Approach: Repository adapters reduce heterogeneous inputs to a shared representation of feature matrix, sample metadata, and resolved site label. Two check families run against it. Modality-agnostic integrity checks cover missingness and its correlation with site, class imbalance, batch, and site effects via cross-validated site-predictability under a permutation null, and leakage from patient-straddle or identifier columns. Oncology completeness checks cover staging, histology, treatment, outcomes, actionable biomarkers, and ICD-O/mCODE conformance. Domain and task knowledge live in declarative YAML packs rather than code, keeping the engine small and extensible. Output is an evidence-linked data card plus Croissant metadata.
Datasets: GDC TCGA-BRCA (clinical, expression, mutations) and one IDC imaging collection; PDC CPTAC proteomics and SEER as stretch targets. Open access tiers throughout.
Jamboree deliverables: A public repository containing the intermediate representation and adapter interface, working batch-effect and leakage checks validated on TCGA-BRCA, a first oncology completeness pack, and a per-patient multimodal linkage matrix spanning GDC and IDC, with a rendered data card.
Tools and environment: Python (pandas, scikit-learn, pydicom), open-tier repository APIs, and modest compute: the design is metadata-first and requires no pixel-level processing.
Team: Assembled and ready (see additional authors), four members spanning computer science and healthcare data analytics, with roles mapped to architecture layers: representation, integrity checks, oncology pack, and adapters.