Présentation : Graph neural networks for predicting patient mortality within one year of hospital admission
Date
2023-10-13
Auteurs
Résumé
Introduction
Predicting long-term patient survival after admission is crucial for improving palliative care and initiating goals of care discussions. Previous research successfully predicted the Hospital-patient One-year Mortality Risk for all causes using routine admission data, but overlooked valuable temporal information from a patient’s multiple hospital visits. Our study proposes to leverage patients’ longitudinal data to predict one-year mortality risk at admission. Our code is publicly available on: https:// github.com/MEDomics-UdeS/POYM
Methods
We analyzed 197,905 hospital visits of 116,002 adult patients admitted to a nonpsychiatric service at the University Hospital of Sherbrooke from July 2011 to June 2021. Two models were compared: a previously developed Random Forest (RF) treating patient visits independently, and a new recurrent neural network (LSTM) model considering longitudinal patient data. 5-folds cross-validation on 82,104 patients admitted between 2011 and 2017 assessed the benefits of temporal analysis. Modeled data types were “AdmDemo” (patient demographics and admission characteristics) and “AdmDemoDx” (also incorporating admission and old comorbidities diagnoses). A final LSTM model was trained on admissions between 2011 and 2017 and tested on a separate holdout set of admissions between 2017 and 2021 (33,898 patients), excluding non-eligible end-of-life care visits.
Results
Our LSTM model demonstrated a raise in performance with an AUROC of 0.93 for “AdmDemoDx” and 0.90 for “AdmDemo”. The DeLong statistical test confirmed significant differences (p-values < 0.05) between the RF model and our approach in both settings. On the holdout set, our model demonstrated AUROC values of 0.89 for “AdmDemoDx” and 0.85 for “AdmDemo”.
Discussion/Conclusion
Our findings demonstrate the effectiveness of temporal analysis in predicting one-year mortality from hospitalization data of different types. Feature importance analysis revealed that simple data available at admission (AdmDemo) are the most discriminant in identifying high-risk patients. Overall, this study emphasizes the significance of leveraging patients’ longitudinal data to advance predictive models and enhance patient care.
Présentation
