Nomogram versus machine learning for predicting synchronous multiple esophageal squamous cell neoplasias using multicenter data

This study is aim to develop and validate a simple, clinical diagnostic tool based on the clinical and pathological characteristics of synchronous multiple esophageal squamous cell neoplasias (ESCNs), thereby assisting clinicians in reducing the rate of missed diagnoses. This retrospective study analyzed the clinical and pathological data of 445 patients with ESCN from three medical centers. Patients were divided into a training group ( n = 349) and an external validation group ( n = 96) based on center of care.We developed a diagnostic nomogram model for synchronous multiple ESCN using logistic regression and Platt scaling calibration. We then compared this model with the best one built using 21 machine learning methods to identify a simpler, more stable, and more accurate model that could assist clinical practice. The incidence of synchronous multiple ESCNs was 10.6% in the training group and 12.5% in the validation group ( P = 0.599). The lesion diameter over than 30 mm (OR = 10.79, 95% CI: 3.86–31.25) and a pathological diagnosis of early invasive carcinoma (EIC) (OR = 6.05, 95% CI: 2.07–20.59) were independent risk factors for synchronous multiple ESCNs. Conversely, non-preference for hot foods (OR = 0.40, 95% CI: 0.18–0.89) was a protective factor. The nomogram model constructed using seven variables had an area under the curve (AUC) of 0.752 (95% CI: 0.617–0.887). Decision curve analysis (DCA) demonstrates clinical net benefit within the threshold probability range of 0.1 to 0.5. In contrast, the MARS model, had an AUC of 0.802 (95% CI: 0.715–0.886). But, its net benefit was limited to a narrow threshold range of 0.07–0.10. Considering clinical utility, stability, and interpretability, the nomogram model is more suitable for predicting synchronous multiple ESCNs. This study developed and validated a nomogram model for predicting synchronous multiple ESCN based on multicenter real-world data. The nomogram model demonstrated good discriminatory power and positive net benefit. While the MARS model showed slightly higher AUC, the nomogram offers practical advantages for clinical application due to its interpretability and ease of use.