Abstract | Comorbidity in patients, along with attendant operations and complications, is associated with reduced long-term survival probability and an increased need for healthcare facilities. This study proposes a user-friendly toolkit to design an adjusted case-mix model of the risk of comorbidity for use by the public for its incremental development. The proposed model, Temporal Comorbidity-Adjusted Risk of Emergency Readmission (T-CARER), introduces a generic method for generating a pool of features from re-categorised and temporal features to create a customised comorbidity risk index. Research on emergency admission has shown that demographics, temporal dimensions, length of stay, and time between admissions can noticeably improve statistical measures related to comorbidities. The model proposed in this study, T-CARER, incorporates temporal aspects, medical procedures, demographics, admission details, and diagnoses. And, it tries to address four weakness areas in popular comorbidity risk indices: robustness, temporal adjustment, population stratication, and inclusion of major associated factors. Three approaches to modelling, a logistic regression, a random forest, and a wide and deep neural network, are designed to predict the comorbidity risk index associated with 30- and 365-day emergency readmissions. The models were trained and tested using England's Hospital Episode Statistics inpatient database for two time-frames: 1999-2004 and 2004-2009, and various risk cut-os. Also, models are compared against implementations of Charlson and Elixhauser's comorbidity indices from multiple aspects. Tests using k fold cross-validation yielded stable and consistent results, with negative mean-squared error variance of -0.7 to -2.9. In terms of c-statistics, the wide and deep neural network and the random forest models outperformed Charlson's and Elixhauser's comorbidity indices. For the 30- and 365-day emergency readmission models, the c-statistics ranged from 0.772 to 0.804 across the timeframes. The wide and deep neural network model generated predictions with high precision, and the random forest model performed better than the regression model, in terms of the micro-average of the F1-score. Our best models yielded precision values in the range of 0.582{0.639, and an average F1-score of 0.730{0.790. The proposed temporal case-mix risk model T-CARER outperforms prevalent models, including Charlson's and Elixhauser's comorbidity indices, with superior precision, F1-score, and c-statistics. The proposed risk index can help monitor the temporal comorbidities of patients and reduce the cost of emergency admissions. |
---|