# Exploratory Algorithms to Aid in Risk of Malignancy Prediction for Indeterminate Pulmonary Nodules

## TL;DR

This study develops a machine learning model combining clinical data and biomarkers to better predict if lung nodules are cancerous, potentially reducing diagnostic delays.

## Contribution

The novel contribution is a machine learning model integrating clinical, radiographic, and biomarker data to predict malignancy in indeterminate pulmonary nodules.

## Key findings

- The model achieved an AUC of 0.872 in training and 0.842 in testing, outperforming the Mayo Score model.
- Key predictors included age, lesion size, pack-years, and biomarkers like hs-CRP, NSE, and CA-125.
- The model could reduce the need for repeated imaging and diagnostic delays in patients with indeterminate nodules.

## Abstract

Annual low-dose CT-based lung cancer screening has been demonstrated to reduce patient mortality relative to chest X-rays, but has significant challenges in distinguishing malignant from non-malignant radiographic findings. To address this clinical need, we applied machine learning to readily accessible clinical information (i.e., patient demographics, clinical characteristics, and radiographic parameters) along with measurements from common circulating biomarkers to develop a method to aid providers with clinical decision making and reduce diagnostic delay inherent to current treatment standards.

Background/Objectives: Lung cancer screening can reduce patient mortality. Multiple issues persist including timely management of patients with a radiologically defined indeterminate pulmonary nodule (IPN), which carries unknown pathological significance. This pilot study focused on combining demographic, clinical, radiographic, and common circulating biomarkers for their ability to aid in IPN risk of malignancy prediction. Methods: A case-control cohort consisting of 379 patients with IPNs (251 stage I lung tumors and 128 nonmalignant nodules) was used for this effort, divided into training (70%) and testing (30%) sets. Demographic variables (age, sex, race, ethnicity), radiographic information (nodule size and location), smoking pack-years, and plasma biomarker levels of CA-125, SCC, CEA, HE4, ProGRP, NSE, Cyfra 21-1, IL-6, PlGF, sFlt-1, hs-CRP, Ferritin, IgG, IgE, IgM, IgA, and Kappa and Lambda Free Light Chains were assessed for this purpose. Results: Multivariable analyses of biomarker, demographic, and radiographic variables yielded a model consisting of age, lesion size, pack-years, history of extrathoracic cancer, upper lobe location, spiculation, hs-CRP, NSE, Ferritin, and CA-125 (AUC = 0.872 in training, 0.842 in testing) with superior performance over the Mayo Score model, which consists of age, lesion size, history of smoking, history of extrathoracic cancer, upper lobe location, and spiculation (AUC = 0.816 in training, 0.787 in testing). Conclusions: In conclusion, a simple reduced algorithm consisting of biomarkers, clinical information, and demographic variables may have value for malignancy prediction of screen-detected IPNs. Upon further validation, this method stands to reduce the need for serial radiographic studies and the risks of diagnostic delay.