Mukul Sehgal, MD
University of South Alabama Children's & Women's Hospital
Mobile, Alabama
Disclosure information not submitted.
Shashank Vadyala, PhD
Researcher, Department of Computational Analysis and modelling
Louisiana Tech University, LA, United States
Disclosure information not submitted.
Amod Amritphale, MD
Assisstant Professor
University of South Alabama, United States
Disclosure information not submitted.
Title: Pediatric Pulmonary hypertension readmissions prediction model using Artificial Intelligence
Introduction: Prediction models are used commonly for predicting outcomes in medicine are mostly based on logistic regression. Most clinical models have lower accuracy, and area under curve (AUC), a measure of a model to predict readmission vs non-readmission in future, due to limited sample size and variable selection. The ability to recognize patterns by AI can be a useful tool in medicine, which can lead to a superior prediction model. We attempt to create a prediction model using AI to help predict 30-day readmissions in pediatric patients with Pulmonary hypertension (PH)
Methods: National Readmission Database (NRD) 2017 was searched for patients < 18 years of age who were diagnosed with PH based on International Classification of Diseases, Tenth Revision (ICD-10). Python® software was used to build a prediction model. Synthetic Minority Oversampling Technique (SMOTE) was used to balance data. Data was divided into training and testing set to evaluate the efficiency of the model.
Results: Of 5.52 million pediatric encounters, 10,501 patients met the selection criteria, with a 14.4% readmission rate, which showed that our data was unbalanced. Based on clinical importance and AI, 14 variables were selected among which length of stay (LOS) during initial admission had highest association with readmissions (F score = 2921), followed by respiratory infection (F=464) and mechanical ventilation (F=297), while use of Nitric oxide was least likely associated with readmissions (F=40). A prediction model was developed using Gradient Boost Classifier (GBC) and SMOTE was found to be the best prediction model (AUC 0.93, accuracy 87%) for predicting the patients who were likely to be readmitted in 30 days. This model was more accurate than traditionally used logistic regression model (AUC 0.65, accuracy 61%).
Conclusion: Compared to traditional methods, machine learning and artificial intelligence has helped us create a superior prediction model, in terms of accuracy and c-statistic, to predict the readmissions among Pediatric PH patients. This study also demonstrates the use of SMOTE technique in data balancing, that can improve the ability to design better prediction models. An external validation of our model using a different data set would be the next step to improve its credibility for clinical use at bedside.