Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2573 - Predictive Modeling of Lung Cancer Subtype with Integration of Clinical and Chest CT Features to Guide Early Brain Metastasis Management
Presenter(s)
Jino Park, MD - University of California Irvine, Orange, CA
J. Park1, P. Buclez2, J. Lubisich3, C. Hui1, J. P. Harris1, and A. B. Simon4; 1Department of Radiation Oncology, University of California - Irvine, Orange, CA, 2California Northstate University College of Medicine, Elk Grove, CA, 3University of California Irvine School of Medicine, Irvine, CA, 4Department of Radiation Oncology, University of California, Irvine, Orange, CA
Purpose/Objective(s):
Optimal treatment of patients presenting with a new diagnosis of lung cancer and synchronous brain metastases remains a clinical challenge. Decisions regarding surgery, radiosurgery, or whole-brain radiation often must be made urgently before complete histologic and molecular information is available. We sought to develop a tool to predict lung cancer subtype using clinical and chest CT features to distinguish subtypes requiring distinct brain metastasis management strategies: (1) small cell lung cancer (SCLC), (2) non–small cell lung cancer harboring EGFR mutation (EGFR[+]), and (3) non–small cell lung cancer without EGFR mutation (EGFR[–]).Materials/Methods:
Records of 303 patients (2016–2025) with metastatic lung cancer at diagnosis with available initial CT chest and MRI brain reports were retrospectively obtained from a single institution. Predefined clinical and imaging features associated with lung cancer subtypes in prior studies were collected. To reduce interpretation bias, a firewalled large language model, blinded to tumor subtype, determined presence or absence of each CT feature. Those without brain metastases were used for training (n=180; EGFR[+]=66, SCLC=57, EGFR[–]=57), and those with brain metastases for testing (n=123; EGFR[+]=53, SCLC=19, EGFR[–]=51). Logistic regression with LASSO (LR) and random forest (RF) models were evaluated using clinical features only (C) and combined clinical plus imaging features (Co).Results:
Model performance metrics including area under the receiver operating characteristic curve (AUC), positive predictive value (PPV), and negative predictive value (NPV) are shown in Table 1. Performance remained good in the testing dataset suggesting that training on those without brain metastases did not overly hinder model performance. In the testing cohort, prediction was superior for EGFR[+] compared with SCLC and EGFR[–]. Integrating imaging features improved performance for EGFR[+] and SCLC with both LR and RF models, but not for EGFR[–]. In one-vs-rest analysis, RF demonstrated comparable performance to LR across subtypes. Important imaging features for EGFR[+] included pleural attachment, miliary pulmonary metastases, fibrosis, and mediastinal lymphadenopathy, while SVC and pleural involvement were associated with SCLC.Conclusion:
Incorporation of clinical and chest CT imaging features into machine learning models can aid discrimination of EGFR[+], SCLC, and EGFR[–] and support earlier multidisciplinary management discussions.
| Subtype | EGFR(+) | SCLC | EGFR(-) | |
| Training | AUC-LR-C | 0.93 | 0.76 | 0.73 |
| AUC-LR-CI | 0.97 | 0.85 | 0.86 | |
| p-value LR (C vs. CI) | <0.01 | <0.01 | <0.01 | |
| AUC-RF-C | 0.98 | 0.94 | 0.94 | |
| AUC-RF-CI | 0.99 | 0.97 | 0.97 | |
| p-value RF (C vs. CI) | 0.01 | 0.01 | 0.01 | |
| Testing | AUC-LR-C | 0.86 | 0.68 | 0.76 |
| AUC-LR-CI | 0.90 | 0.78 | 0.80 | |
| p-value LR (C vs. CI) | 0.02 | 0.02 | 0.13 | |
| AUC-RF-C | 0.89 | 0.67 | 0.78 | |
| AUC-RF-CI | 0.91 | 0.77 | 0.81 | |
| p-value RF (C vs. CI) | 0.03 | 0.01 | 0.20 | |
| p-value CI (LR vs. RF) | 0.37 | 0.56 | 0.62 | |
| PPV-LR-CI | 0.84 | 0.33 | 0.68 | |
| NPV-LR-CI | 0.86 | 0.89 | 0.69 |