468 Background: Despite evidence supporting early palliative care integration, utilization remains suboptimal with only 13.1% of cancer patients receiving services. We developed and validated a deep learning clinical decision support system (CDSS) to identify cancer patients who would benefit from palliative care (PC) consultation. Methods: Using National Inpatient Sample data (2018-2021, n=734,212), we trained a multi-output deep learning model to predict PC needs and in-hospital mortality for patients with lung, breast, prostate, and colon cancers. The model incorporated 23 features including demographics, diagnoses (ICD-10), comorbidities, and severity indices. We employed focal loss functions and moderate SMOTE balancing to address class imbalance. Model performance was evaluated using balanced accuracy, sensitivity, specificity, and F1-scores. External validation was performed on 2022 data (n=181,636). Optimal thresholds were determined using ROC analysis to maximize clinical utility. Results: The model achieved balanced accuracy of 80.2% for palliative care prediction and 80.1% for mortality prediction. For PC identification, sensitivity was 80% with specificity of 80% (F1-score=0.51). Mortality prediction showed 81% sensitivity and 80% specificity (F1-score=0.34). External validation on 2022 data demonstrated consistent performance (PC: sensitivity 80%, precision 41%; mortality: sensitivity 81%, precision 23%). The model identified key predictors including metastasis (18.8% bone metastasis prevalence), severity index (mean 5.27), and age-comorbidity interactions. Prostate cancer showed lowest PC utilization (8.4%) despite substantial metastatic disease (35.4%). Conclusions: This validated CDSS demonstrates robust performance in identifying cancer patients who would benefit from palliative care, with high sensitivity ensuring few missed cases. The tool addresses critical quality gaps in palliative care delivery and could facilitate timely referrals. Implementation could improve care quality while optimizing resources. Next steps include prospective clinical validation and integration into EMR for real-time decision support. Deep learning model performance for palliative care and mortality prediction. Outcome Dataset Sensitivity Specificity Precision F1-Score AUC Balanced Accuracy Palliative Care Training (2018-21) 587,370 0.802 0.798 0.370 0.507 0.876 0.800 Test (2018-21) 146,843 0.802 0.768 0.370 0.506 0.872 0.785 Validation (2022) 181,636 0.799 0.806 0.410 0.542 0.878* 0.802 In-Hospital Mortality Training (2018-21) 587,370 0.805 0.802 0.213 0.337 0.882 0.804 Test (2018-21) 146,843 0.808 0.802 0.213 0.337 0.879 0.805 Validation (2022) 181,636 0.806 0.795 0.234 0.363 0.881* 0.801 *AUC estimated from balanced accuracy and threshold optimization.
Amalraj et al. (Wed,) studied this question.