A machine-learning-based prediction of non-home discharge among acute heart failure patients

Abstract

Background: Scarce data on factors related to discharge disposition in patients hospitalized for acute heart failure (AHF) were available, and we sought to develop a parsimonious and simple predictive model for non-home discharge via machine learning.

Methods: This observational cohort study using a Japanese national database included 128,068 patients admitted from home for AHF between April 2014 and March 2018. The candidate predictors for non-home discharge were patient demographics, comorbidities, and treatment performed within 2 days after hospital admission. We used 80% of the population to develop a model using all 26 candidate variables and using the variable selected by 1 standard-error rule of Lasso regression, which enhances interpretability, and 20% to validate the predictive ability.

Results: We analyzed 128,068 patients, and 22,330 patients were not discharged to home; 7,879 underwent in-hospital death and 14,451 were transferred to other facilities. The machine-learning-based model consisted of 11 predictors, showing a discrimination ability comparable to that using all the 26 variables (c-statistic: 0.760 [95% confidence interval, 0.752-0.767] vs. 0.761 [95% confidence interval, 0.753-0.769]). The common 1SE-selected variables identified throughout all analyses were low scores in activities of daily living, advanced age, absence of hypertension, impaired consciousness, failure to initiate enteral alimentation within 2 days and low body weight.

Conclusions: The developed machine learning model using 11 predictors had a good predictive ability to identify patients at high risk for non-home discharge. Our findings would contribute to the effective care coordination in this era when HF is rapidly increasing in prevalence.

Keywords: Claims database analysis; Clinical epidemiology; Heart failure; Machine learning; Non-home discharge.

Fig. 1

Flowchart of patient selection. DPC Diagnosis procedure combination

Fig. 2

Cross-validation plot and coefficient paths of variable selection. (A) Cross-validation plot of mean squared error corresponding to smoothing parameter λ with standard errors. (B) Coefficient paths of variable selection. Legends show the selected predictors selected at the point of λ_SE in the order of standardized coefficient values. λ_SE, the largest λ among λ for which the cross-validation function is within one standard error of the minimum of the cross-validation function (λ = 0.0100); λ_CV, λ where the cross-validation function is minimum (λ = 0.000152); NYHA New York Heart Association

Fig. 3

Variable importance of the variables selected by 1 standard-error rule of lasso regression. SE standard-error, NYHA New York Heart Association

Fig. 4

Receiver operating characteristic curve and calibration in primary analysis. (A) Receiver operating characteristic curves of the model using 1SE-selected variables and the model using all variables. (B) Calibration plots showing the models predicting non-home discharge among patients admitted for acute heart failure using the 1SE-selected variables and the model using all variables. SE standard-error, E:O estimated to observed ratio, CITL calibration in the large, AUC area under receiver operating curve, CI confidence interval