# Deep Learning-Based 30-Day Mortality Prediction in Critically Ill Bone and Bone Marrow Metastasis Patients: A Multicenter Retrospective Cohort Study

## TL;DR

A deep learning model called TabNet accurately predicts 30-day mortality in critically ill cancer patients with bone and bone marrow metastases, helping guide treatment decisions.

## Contribution

TabNet outperformed existing models in predicting mortality and was validated across diverse patient cohorts.

## Key findings

- TabNet achieved high accuracy (AUC 0.878) in predicting 30-day mortality in critically ill bone metastasis patients.
- The model maintained strong performance in both same-region and cross-regional validations.
- Key predictors included SOFA score, serum calcium, and albumin, identified through interpretability analyses.

## Abstract

Patients with cancer that spreads to the bones and bone marrow often become critically ill and face a high risk of death, which makes reliable prognostic tools essential for determining whether intensive care should prioritize aggressive treatment or hospice approaches, yet existing scoring systems remain limited. This multicenter study used data from hospitals in the United States and China, incorporating routinely collected clinical information from the first day of intensive care to construct advanced computer models for predicting 30-day survival, and demonstrated that TabNet provided the most accurate and consistent performance across diverse cohorts, with its deployment as an online calculator enabling early and transparent risk stratification that supports timely clinical decision-making. These findings highlight the potential of deep-learning model to improve prognostic evaluation in oncologic critical care and to inform future research directions and policy strategies aimed at optimizing outcomes for patients with advanced cancer.

Bone and bone marrow Metastasis (BBM) are life-threatening complications of advanced malignancies, frequently requiring intensive care and associated with high short-term mortality. However, prognostic tools specifically tailored to critically ill BBM patients are limited. This multicenter cohort study aimed to develop and validate deep learning models for predicting 30-day mortality using ICU data from MIMIC-IV, eICU-CRD, and the First Affiliated Hospital of Xinjiang Medical University. After univariate screening, XGBoost-Boruta and Lasso regression identified 11 key clinical features within 24 h of ICU admission. Thirteen deep learning models were trained using five-fold cross-validation, and their performance was evaluated through AUC, average precision, calibration, and decision curves. TabNet achieved the best internal performance (AUC 0.878; AP 0.940) and maintained strong discrimination in both same-region (eICU: AUC 0.840; AP 0.932) and cross-regional (Xinjiang: AUC 0.831; Accuracy 80.5%) validation. SHAP and attention-based interpretability analyses consistently identified SOFA, serum calcium, and albumin as dominant predictors. A TabNet-based online calculator was subsequently deployed to enable bedside mortality risk estimation. In conclusion, TabNet demonstrates potential as an accurate and interpretable tool for early mortality risk stratification in critically ill BBM patients, offering support for more timely and individualized decision-making in BBM-related critical care.