# Differentiation of early-stage tumors from benign lesions manifesting as pure ground-glass nodule: a clinical prediction study based on AI-derived quantitative parameters

## TL;DR

This study uses AI-derived parameters to help doctors distinguish between benign and early-stage tumor nodules in the lungs, improving clinical decision-making.

## Contribution

A novel clinical prediction nomogram is developed using AI-derived quantitative parameters to differentiate benign from malignant pure ground-glass nodules.

## Key findings

- A nomogram with nine predictors achieved AUCs of 0.696 in training and 0.672 in external validation.
- Calibration and decision curve analysis confirmed the model's clinical utility and net benefit.
- The nomogram provides a visual score for individualized patient management decisions.

## Abstract

Differentiating between benign and malignant pure ground-glass nodule (pGGN) is of great clinical significance. The aim of our study was to evaluate whether AI-derived quantitative parameters could predict benignity versus early-stage tumors manifesting as pGGN.

A total of 1,538 patients with pGGN detected by chest CT at different campuses of our hospital from May 2013 to December 2023 were retrospectively analyzed. This included CT and clinical data, as well as AI-derived quantitative parameters. All patients were randomly divided into a training group (n=893), an internal validation group (n=382), and an external validation group (n=263). Hazard factors for early-stage tumors were identified using univariate analysis and multivariate logistic regression analysis. Independent risk factors were then screened, and a prediction nomogram was constructed to maximize predictive efficacy and clinical application value. The performance of the nomogram was evaluated using ROC curves and calibration curves, while decision curve analysis (DCA) was used to assess the net benefit prediction threshold.

The final logistic model included nine independent predictors (age, location, minimum CT value, standard deviation, kurtosis, compactness, energy, costopleural distance, and volume) and was developed into a user-friendly nomogram. The AUCs of the ROC curves in the training, internal validation, and external validation cohorts were 0.696 (95% CI: 0.638–0.754), 0.627 (95% CI: 0.533–0.722), and 0.672 (95% CI: 0.543–0.801), respectively. The calibration plot demonstrated a good correlation between observed and predicted values, and the nomogram remained valid in the validation cohort. DCA showed that the model’s predictive performance was acceptable, providing substantial net benefit for clinical application.

The clinical prediction nomogram, based on AI-derived quantitative parameters, visually displays an overall score to differentiate benign lesions from early-stage tumors manifesting as pGGN. This nomogram may serve as a convenient screening tool for clinical use and provides a reference for formulating individualized follow-up and treatment plans for patients with pGGN.