Michael Brudno

Machine learning models have seen a lot of success in medical image segmentation domain. However, one of the challenges that they face are confounders or shortcuts: spurious correlations or biases in the training data that affect the resulting models. One example of such confounders for surgical machine learning is the setup of surgical equipment, including tools and lighting. Using the task of identification of safe and dangerous zones of dissection in laparoscopic cholecystectomy images and videos as a use-case, we inspect two equipment-induced biases: the location of surgical tools in the field of view and the direction of lighting, both of which are tightly related to the region of interest. We propose methods for evaluating the severity of these biases based on measurements of model consistency under real or simulated exposure to these shortcuts, and augmentation-based methods for mitigating them. We show that our tool bias mitigations based on pasting tool images in random locations improve the models’ prediction consistency under tool movements by 9 percentage points in the most inconsistent cases, and by 4 percentage points on average. Our lighting bias mitigations based on simulated lighting adjustments help reduce fraction of pixels originally predicted as belonging to the dangerous zone that may flip to safe under light changes from 5% to 1.5%, without compromising segmentation quality.

GDM domain schema. The GDM schema summarizes the full set of data elements collected to represent the journey of a person living with cancer by integrating (i) foundational structured data encoded in routine clinical systems using standard terminologies (Fig. 2), (ii) high-value variables that often require manual abstraction from unstructured documentation, and (iii) derived variables calculated from (i) and (ii). The schema organizes elements into domains centered on the patient with cancer. The person-centered domain includes demographics, geographical linkage to area SDoH, individual-level SDoH (education, employment, marital status), hospitalization events, comorbidities, vital signs/performance status, and laboratory testing. The treatment-centered domain includes surgeries (procedures and pathology reports) and radiotherapy (course/phase, technique/modality, and dosimetry), systemic therapies (drug-level exposure with start/end dates and modification/discontinuation reasons), and clinical trial participation. The diagnosis-centered domain includes cancer diagnosis (ICD-O-3 tumor classification coding; diagnostic confirmation framework), cancer staging [grade, behavior, morphology, tumor–node–metastasis (TNM)], and genomic profiling of tumors (somatic alterations and sequencing panel metadata). Lastly, the outcomes-centered domain includes imaging (event dates, imaging types, radiologist assessment of cancer status and response) and clinical visits (event dates, visit types, and oncologist assessment of cancer status and response). The outcomes-centered domain also harmonizes temporally anchored tumor-centric assessments and person-centric measures to compute real-world endpoints, including overall survival (OS), progression-free survival (PFS), time to treatment discontinuation (TTD), time to next treatment (TTNT), and real-world response rate (rwRR). Together, these domains represent the complete GDM scope and illustrate how combining coded structured data with curated abstraction enables robust longitudinal phenotyping for real-world evidence generation. AE, adverse event; CNS, central nervous system; CTCAE, Common Terminology Criteria for Adverse Events; DOB, Date of Birth; ECOG, Eastern Cooperative Oncology Group; EOD, extent of disease; KPS, Karnofsky performance status; LTFU, Lost to Follow-Up; MGMT, O6-methylguanine-DNA methyltransferase; NDI, National Death Index; SDV, Source Data Verification; SSDI, Site-Specific Data Item. (Icon in this figure adapted from an image generated by ChatGPT, February 6, 2026, OpenAI, https://chat.openai.com.)

GDM interoperability. The figure illustrates how the GDM bridges standardized point-in-time oncology data exchange and population-level analytics by integrating and extending existing interoperability frameworks. (Left) The mCODE provides standardized HL7 FHIR profiles for discrete oncology concepts, including MedicationRequest (e.g., chemotherapy agents), Procedure (e.g., biopsy), and Observation (e.g., disease status snapshots). Although mCODE enables standardized exchange of individual clinical events, it has limited native support for longitudinal patterning and multiagent treatment constructs. (Middle) The GDM’s two-layer architecture. Layer A (abstracted foundation) captures granular, source-level clinical data aligned to mCODE-compatible FHIR profiles while preserving provenance and clinical context. This layer includes agent-level therapy exposure (discrete drugs with start and end dates and dose information), person-centric clinical events (e.g., hospitalizations and treatment discontinuation reasons), and tumor-centric assessments (e.g., imaging and clinician documentation). Building on this foundation, layer B (derived intelligence) applies model-defined derivation rules and algorithms to synthesize these granular inputs into longitudinally, analytically meaningful constructs. These include multiagent regimens, lines of therapy, and algorithmically derived endpoints such as progression-free survival and overall survival, with traceable logic linking derived outputs back to underlying source data. (Right) GDM outputs are transformed through ETL processes to support interoperability with standardized vocabularies and episode-based presentations compatible with data models like OMOP CDM. This enables population-level analytics, cohort generation, and downstream research across institutions. Together, this workflow supports the collection of heterogeneous data from a wide range of sources while maintaining transparent provenance, extensibility, and alignment with existing oncology data standards. (This figure adapted from an image generated by ChatGPT, February 6, 2026, OpenAI, https://chat.openai.com.)

Adequately powered analyses in precision oncology often require combining cohorts across institutions. However, integration is constrained by the least granular source and may become infeasible when data elements are too heterogeneous to harmonize and map to a common data model. This challenge is acute in multi-institutional precision oncology research, in which real-world evidence requires harmonized clinico-omic data integration. Existing models often lack sufficient treatment patterns, outcomes, and genomic data, limiting interoperability and scalability. To address these gaps, American Association for Cancer Research Project Genomics Evidence Neoplasia Information Exchange (GENIE) developed the GENIE Data Model (GDM), a comprehensive, open-source oncology data model for scalable, consistent, and interoperable data collection across solid tumors, designed to effectively capture the journey of patients with cancer. Through iterative consensus-building, four working groups comprising 13 subject matter experts defined data elements across multiple clinical domains: patient characteristics, imaging, diagnosis, surgery, histopathology, biomarkers, systemic therapy, radiation, clinical trial history, disease response and outcomes, and social determinants of health. Elements were defined using standardized terminologies and permissible values to support mapping to Health Level 7 Fast Healthcare Interoperability Resources, Observational Medical Outcomes Partnership, and other existing oncology standards. The model architecture distinguishes manually abstracted elements from computationally collected elements, enabling parallel workflows. The GDM provides an extensible framework that addresses critical gaps and enables scalable, harmonized data collection essential for precision oncology and real-world evidence generation. Significance:The GDM is an open-source, ready-to-implement, solid tumor framework that standardizes clinical and genomic data capture across institutions to effectively capture the journey of patients with cancer. It enables interoperable mapping to existing standards, accelerates multisite study start-up, and supports scalable real-world evidence generation in precision oncology.

Detailed information on GDM Logical Model

Detailed information on clinical domains

GDM logical model diagram. The GDM logical model is shown as a Unified Modeling Language diagram demonstrating relationships between different entities. A, Full schema view; the blue box shows foundational entities (data_partner, organization, cohort, patient, death, ehr_artifact) linked to clinical entities that represent oncology care (medication, procedure, visit, report, specimen, treatment, pathology results). Relationship cardinalities and linkages (e.g., orderedBy, performedBy) define dependencies across entities. B, (Blue box) Expanded view of the report entity illustrating selected attributes (e.g., patientReference, date, categoryConcept, providerReference, specimenReference, visitReference, and identifier). The model aligns with HL7 FHIR resource structures, and attributes are represented with unique model_id paths (e.g., report∼categoryConcept) to support both manual abstraction and computational derivation.

Overview of existing models and the GDM.

Overview of GDM domains + data elements.

Adequately powered analyses in precision oncology often require combining cohorts across institutions. However, integration is constrained by the least granular source and may become infeasible when data elements are too heterogeneous to harmonize and map to a common data model. This challenge is acute in multi-institutional precision oncology research, in which real-world evidence requires harmonized clinico-omic data integration. Existing models often lack sufficient treatment patterns, outcomes, and genomic data, limiting interoperability and scalability. To address these gaps, American Association for Cancer Research Project Genomics Evidence Neoplasia Information Exchange (GENIE) developed the GENIE Data Model (GDM), a comprehensive, open-source oncology data model for scalable, consistent, and interoperable data collection across solid tumors, designed to effectively capture the journey of patients with cancer. Through iterative consensus-building, four working groups comprising 13 subject matter experts defined data elements across multiple clinical domains: patient characteristics, imaging, diagnosis, surgery, histopathology, biomarkers, systemic therapy, radiation, clinical trial history, disease response and outcomes, and social determinants of health. Elements were defined using standardized terminologies and permissible values to support mapping to Health Level 7 Fast Healthcare Interoperability Resources, Observational Medical Outcomes Partnership, and other existing oncology standards. The model architecture distinguishes manually abstracted elements from computationally collected elements, enabling parallel workflows. The GDM provides an extensible framework that addresses critical gaps and enables scalable, harmonized data collection essential for precision oncology and real-world evidence generation. Significance: The GDM is an open-source, ready-to-implement, solid tumor framework that standardizes clinical and genomic data capture across institutions to effectively capture the journey of patients with cancer. It enables interoperable mapping to existing standards, accelerates multisite study start-up, and supports scalable real-world evidence generation in precision oncology.