Interpretable machine learning for prediction of leptomeningeal metastasis in lung adenocarcinoma: a multi-centre retrospective study via "Prompt" model.

BACKGROUND: Leptomeningeal metastasis (LM) is a serious complication of lung adenocarcinoma (LUAD). Owing to the difficulty in early diagnosis and the limited clinical benefit of available treatments, affected patients experience extremely poor outcomes. Thus, the machine-learning-enhanced predictive framework Prompt was developed to optimise early clinical identification and prognostic assessment. This model was designed to determine major risk factors for LM in patients with LUAD, estimate the possibility of LM onset and predict survival, thereby supporting clinical decision-making via an early-warning system and accurate survival prediction. METHODS: Clinical data of patients diagnosed with LUAD complicated by LM (LUAD-LM) between 1 January 2012 and 1 January 2025 were obtained from electronic medical databases across multiple medical centres. Embedded feature selection was conducted using tree-based feature importance ranking. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) to determine the most suitable machine-learning model. Shapley additive explanation (SHAP) values were applied to interpret the high-risk diagnostic model, and survival outcomes were examined using Cox proportional hazards regression. RESULTS: Among 6,784 screened LUAD cases, 181 individuals met the inclusion criteria for patients with confirmed LUAD-LM. A high-risk diagnostic model for LUAD-LM was developed using the Tabular Prior-data Fitted Network (TabPFN), an approach recently published in Nature. In the test cohort, the model achieved an AUC of 0.842, with an accuracy of 0.880, a precision of 0.900, a recall of 0.947 and an F1 score of 0.923, reflecting consistently strong discriminatory ability. SHAP value analysis revealed that total CO2 on arterial blood gas, TTLS, indirect bilirubin, EGFR exon 21 L858R mutation status, Karnofsky Performance Status (KPS) score and serum Ca2+ levels were key contributors to LUAD-LM. For prognostic assessment, KPS score, serum Na+ levels and serum Cl- concentrations were crucial determinants. A web-based calculator built on the Streamlit framework facilitated clinical application. CONCLUSION: Prompt framework delivers exploratory LM risk and survival estimations for Chinese patients with LUAD, serving as a preliminary risk-stratification reference rather than a universally applicable clinical diagnostic tool pending prospective external validation across diverse populations and healthcare systems.