Toward Scalable Early Cancer Detection: Evaluating EHR-Based Predictive Models Against Traditional Screening...

Area of Science:

• Oncology
• Medical Informatics
• Data Science

Background:

• Current cancer screening guidelines use limited criteria like age or smoking history.
• Electronic Health Records (EHRs) offer extensive longitudinal patient data for improved risk prediction.
• Limited evidence exists comparing EHR-based models with traditional risk factors for cancer screening.

Purpose of the Study:

• To evaluate the clinical utility of EHR-based predictive models against traditional risk factors for identifying individuals at high risk for cancer.
• To compare the effectiveness of baseline EHR models and advanced EHR foundation models in cancer risk stratification.

Main Methods:

• Systematic evaluation of EHR-based predictive models versus traditional risk factors (gene mutations, family history).
• Utilized data from the All of Us Research Program (>865,000 participants) integrating EHR, genomic, and survey data.
• Assessed predictive performance across eight major cancer types and an expanded set of 26 cancer types using an EHR foundation model.

Main Results:

• Baseline EHR-based models showed a 3- to 6-fold higher enrichment of true cancer cases compared to traditional risk factors alone.
• EHR models demonstrated superior performance whether used independently or complementarily with traditional factors.
• The advanced EHR foundation model further enhanced predictive accuracy across 26 cancer types.

Conclusions:

• EHR-based predictive modeling offers a more effective strategy for identifying high-risk individuals for cancer screening.
• Foundation models trained on comprehensive patient data show significant potential for precise and scalable early cancer detection.
• Integrating EHR data into predictive models can overcome limitations of current screening guidelines.