Predicting Diabetic Readmission Rates: Moving Beyond Hba1c

Abstract

Hospital readmission is considered an effective measurement of care provided within healthcare. Being able to risk identify patients facing a high likelihood of unplanned hospital readmission in the next 30-days could allow for further investigation and possibly prevent the readmission. Current models, such as LACE, sacrifice accuracy in order to allow for end-users to have a straight forward and simple experience. This study acknowledges that while HbA1c is important, it may not be critical in predicting readmissions. It also investigates the hypothesis that using machine learning on a wide feature, making use of model diversity, and blending prediction will improve the accuracy of readmission risk predictions compared with existing techniques. A dataset originally containing 100,000 admissions and 56 features was used to evaluate the hypothesis. The results from the study are encouraging and can help healthcare providers improve inpatient diabetic care.

Predicting 30-Day Hospital Readmission for Diabetes Patients using Multilayer Perceptron

“…Hence clustering was performed to group similar observations into the same group (cluster). As this study focuses on improving the current classification models and not bringing up novel pre-processing techniques, this clustering step will mainly follow the scheme from existing literature [11], [14].…”

“…However, the data used is old compared to the other models and there might be possible improvements in the factor selection process as 33 out of 56 variables were used for analysis. In [11], through Machine Learning algorithm achieved 0.70 -079 c-statistics respectively for 30-70 and 0-30 age group on the same dataset with a sensitivity of 43.63% -49.78% and specificity of 82.62% -89.19%. Compare to the two previous models, the researcher applied a different algorithm to each defined population segments.…”

“…Through initial research, it appeared that the LACE index (Length of stay, Acuity of admission, Charlson comorbidity index and Emergency visits) is so far the most preferred model of readmission prediction due to its ease use and implementation by a healthcare professional [10]- [13]. Despite its assets, this index does not pass the acceptable threshold achieving 0.56 to 0.63 c-statistics for unplanned short-term diabetics readmission prediction, hence urging the need for further research to improve the model's performance [4], [11]. To cope with the limitations of the previous approach this review will only list prediction models for unplanned readmission for diabetics limited to acceptable and above c-statistics.…”

“…As real-world medical data are often noisy, a particular focus will be led on pre-processing task handling both missing data and inconsistencies but also by reducing the dataset and optimizing it for further model deployment [19]. While some pre-processing steps are based on an understanding of the data and background, this study will combine and implement the most relevant pre-processing identified in the body of literature [11], [14], [15]. As such, Hotdeck imputation with Approximate Bayesian Bootstrap (ABB), ICD-9-DM clustering and Random Forest feature selection constitutes some of the key pre-processing tasks.…”

Enhanced Predictive Modelling for 30-Day Readmission Diabetes Patients Based on Data Normalization Analysis

“…In [9], several different classifications were proposed, as the study divided patients into 3 groups according to Age [<30, between [30 -70], > 70]. A separate model was built for each group using some ML algorithms or combining them such as (random forest, different types of gradient enhanced trees, and SVM).…”