Scholarlink

Background <p>Colorectal cancer (CRC) is the fourth most common and second deadliest cancer in the US. Screening is effective at reducing CRC incidence and mortality, but rates of screening remain suboptimal. There are no sensitive machine learning models for accurately identifying individuals at risk for colorectal cancer or precancerous polyps.</p> Objectives <p>The aim of our study was to develop and validate a novel machine learning model that uses multimodal Electronic Health Record (EHR) data, including the most recent complete blood count (CBC), basic metabolic panel (BMP), ICD codes, and medications, to estimate a patient’s likelihood of having CRC or an advanced precancerous lesion.</p> Methods <p>We developed ColAI, an L1-regularized logistic regression model trained on 1-year trailing EHR data, to predict CRC or advanced adenoma at screening colonoscopy. Labs are treated as continuous variables, while ICD codes and medications are represented as binary indicators of presence. ColAI was trained using 87,825 screening colonoscopies and validated using 21,957 independent colonoscopies between August 1, 2020, and March 31, 2024, from the NYU Langone Health system.</p> Results <p>ColAI achieved an AUROC of 0.93 for CRC and 0.98 for CRC or advanced adenoma. Performance remained consistent across different hospitals and time periods within NYU Langone, demonstrating strong generalizability. Performance also remained consistent between first and follow-up colonoscopies, decreasing concern for selection bias.</p> Conclusions <p>ColAI accurately identifies patients at elevated risk for CRC using only routine EHR data. It has the potential to enhance targeted outreach to high-risk, unscreened individuals and improve early cancer detection at the population level.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Identifying Patients at High Risk for Colorectal Carcinoma Using the Electronic Health Record

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Yuri Ahuja,
  • Xucong Meng,
  • Aasma Shaukat

摘要

Background

Colorectal cancer (CRC) is the fourth most common and second deadliest cancer in the US. Screening is effective at reducing CRC incidence and mortality, but rates of screening remain suboptimal. There are no sensitive machine learning models for accurately identifying individuals at risk for colorectal cancer or precancerous polyps.

Objectives

The aim of our study was to develop and validate a novel machine learning model that uses multimodal Electronic Health Record (EHR) data, including the most recent complete blood count (CBC), basic metabolic panel (BMP), ICD codes, and medications, to estimate a patient’s likelihood of having CRC or an advanced precancerous lesion.

Methods

We developed ColAI, an L1-regularized logistic regression model trained on 1-year trailing EHR data, to predict CRC or advanced adenoma at screening colonoscopy. Labs are treated as continuous variables, while ICD codes and medications are represented as binary indicators of presence. ColAI was trained using 87,825 screening colonoscopies and validated using 21,957 independent colonoscopies between August 1, 2020, and March 31, 2024, from the NYU Langone Health system.

Results

ColAI achieved an AUROC of 0.93 for CRC and 0.98 for CRC or advanced adenoma. Performance remained consistent across different hospitals and time periods within NYU Langone, demonstrating strong generalizability. Performance also remained consistent between first and follow-up colonoscopies, decreasing concern for selection bias.

Conclusions

ColAI accurately identifies patients at elevated risk for CRC using only routine EHR data. It has the potential to enhance targeted outreach to high-risk, unscreened individuals and improve early cancer detection at the population level.