RESUMO
A hospital readmission risk prediction tool for patients with diabetes based on electronic health record (EHR) data is needed. The optimal modeling approach, however, is unclear. In 2,836,569 encounters of 36,641 diabetes patients, deep learning (DL) long short-term memory (LSTM) models predicting unplanned, all-cause, 30-day readmission were developed and compared to several traditional models. Models used EHR data defined by a Common Data Model. The LSTM model Area Under the Receiver Operating Characteristic Curve (AUROC) was significantly greater than that of the next best traditional model [LSTM 0.79 vs Random Forest (RF) 0.72, p<0.0001]. Experiments showed that performance of the LSTM models increased as prior encounter number increased up to 30 encounters. An LSTM model with 16 selected laboratory tests yielded equivalent performance to a model with all 981 laboratory tests. This new DL model may provide the basis for a more useful readmission risk prediction tool for diabetes patients.
Assuntos
Aprendizado Profundo , Diabetes Mellitus , Humanos , Readmissão do Paciente , Memória de Curto Prazo , Curva ROCRESUMO
OBJECTIVE: In the ACCORD trial, intensive treatment of patients with type 2 diabetes and high cardiovascular (CV) risk was associated with higher all-cause and CV mortality. Post hoc analyses have failed to implicate rapid reduction of glucose, hypoglycemia, or specific drugs as the causes of this finding. We hypothesized that exposure to injected insulin was quantitatively associated with increased CV mortality. RESEARCH DESIGN AND METHODS: We examined insulin exposure data from 10,163 participants with a mean follow-up of 5 years. Using Cox proportional hazards models, we explored associations between CV mortality and total, basal, and prandial insulin dose over time, adjusting for both baseline and on-treatment covariates including randomized intervention assignment. RESULTS: More participants allocated to intensive treatment (79%) than standard treatment (62%) were ever prescribed insulin in ACCORD, with a higher mean updated total daily dose (0.41 vs. 0.30 units/kg) (P < 0.001). Before adjustment for covariates, higher insulin dose was associated with increased risk of CV death (hazard ratios [HRs] per 1 unit/kg/day 1.83 [1.45, 2.31], 2.29 [1.62, 3.23], and 3.36 [2.00, 5.66] for total, basal, and prandial insulin, respectively). However, after adjustment for baseline covariates, no significant association of insulin dose with CV death remained. Moreover, further adjustment for severe hypoglycemia, weight change, attained A1C, and randomized treatment assignment did not materially alter this observation. CONCLUSIONS: These analyses provide no support for the hypothesis that insulin dose contributed to CV mortality in ACCORD.