1240 Predicting lung cancer (LC) diagnosis in the United States (US) real-world setting using an electronic medical record (EMR) database and supervised machine learning
Year: 2024
Type: conference-abstract
Abstract: Background Annual LC screening with low-dose computed tomography is recommended for people aged 50–80 years with 20 pack-year smoking history who currently smoke or have quit within 15 years. Screening leads to LC diagnosis in earlier stages, associated with better survival, when patients may benefit more from therapy, especially with recent approvals for early-stage NSCLC of neoadjuvant (nivolumab), perioperative (pembrolizumab), and adjuvant (pembrolizumab, atezolizumab) immunotherapy. Smoking status is often captured in EMRs, whereas details of pack-years and quitting dates are frequently missing. This study aimed to develop an explainable risk prediction model for LC diagnosis using a real-world database. Methods Using the US IQVIA ambulatory EMR database (~94 million patient records), we included adults aged 50–80 years with outpatient visit (index date) in 2014–2019 and ≥365 days of database observation pre-index date, excluding those with no post-index data, any prior cancer diagnosis, or records suggesting prior LC diagnosis. Patient characteristics and risk factors, prespecified based on known/suspected association with LC, that were associated in univariate logistic regression with LC diagnosis at 1, 2, or 3 years post-index were assessed in multivariable models, developed using supervised machine learning. LC diagnosis was defined as (A) ≥2 ICD-9/10 LC diagnosis codes ≥60 days apart or (B) Definition A plus post-index biopsy/bronchoscopy. We tested the model among only nonsmoking patients and developed a second model assessing screening guidelines-defined factors (age, smoking status, pack-years). Model performances were evaluated using area under the curve (AUC). Database cutoff was mid-2023. Results The 8,152,918 eligible patients were randomly assigned (80:20) to training (n=6,523,573) and testing datasets (n=1,629,345). Mean (SD) age was 63 (9) years, 58% were women, 75%/7%/18% White/Black/All other races, 29% current/former smokers and 37% nonsmokers (34% unknown smoking status). Factors important in univariate analyses informing the multivariable model are listed in table 1. Post-index 1/2/3 years, 0.10%/0.15%/0.19% of patients, respectively, were observed to have LC per Definition A, and, each year, 0.01% per Definition B. (table 2) reports model AUCs. Conclusions The main model without pack-years performed better than a base model with only screening guidelines-defined factors -- highlighting risk factors beyond age and smoking, such as inflammatory lung diseases, LC family history, substance abuse, major depression. Findings demonstrate feasibility of using EMR data without pack-years to identify LC risk for shared decision-making regarding LC screening, including among nonsmokers. Future work will assess model performance with unsupervised learning methods and information on medications and laboratory values. Acknowledgements We thank Yu (Julia) Feng for programming support. Medical writing and editorial assistance were provided by Elizabeth V. Hillyer, DVM (freelance). This assistance was funded by Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA. Ethics Approval This noninterventional study using deidentified data was exempt from institutional review board approval.
Authors: Melissa L Santorelli, Fangyu Li, Hozefa A Divan
Institution: Merck & Co., Inc., Rahway, NJ, USA (United States)
Language: English
Subfield: Pulmonary and Respiratory Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W4404065274