Prediction of Clinical Deterioration in Hospitalized Adult Patients with Hematologic Malignancies Using a Neural Network Model

Abstract

Introduction: Clinical deterioration (ICU transfer and cardiac arrest) occurs during approximately 5-10% of hospital admissions. Existing prediction models have a high false positive rate, leading to multiple false alarms and alarm fatigue. We used routine vital signs and laboratory values obtained from the electronic medical record (EMR) along with a machine learning algorithm called a neural network to develop a prediction model that would increase the predictive accuracy and decrease false alarm rates.

Design: Retrospective cohort study.

Setting: The hematologic malignancy unit in an academic medical center in the United States.

Patient population: Adult patients admitted to the hematologic malignancy unit from 2009 to 2010.

Intervention: None.

Measurements and main results: Vital signs and laboratory values were obtained from the electronic medical record system and then used as predictors (features). A neural network was used to build a model to predict clinical deterioration events (ICU transfer and cardiac arrest). The performance of the neural network model was compared to the VitalPac Early Warning Score (ViEWS). Five hundred sixty five consecutive total admissions were available with 43 admissions resulting in clinical deterioration. Using simulation, the neural network outperformed the ViEWS model with a positive predictive value of 82% compared to 24%, respectively.

Conclusion: We developed and tested a neural network-based prediction model for clinical deterioration in patients hospitalized in the hematologic malignancy unit. Our neural network model outperformed an existing model, substantially increasing the positive predictive value, allowing the clinician to be confident in the alarm raised. This system can be readily implemented in a real-time fashion in existing EMR systems.

Conflict of interest statement

Competing Interests: The authors have declared that no competing interests exist.

Figures

Fig 1. Representative neural network model demonstrating a simplified version of the neural network used to predict clinical deterioration in hematologic malignancy patients.

The features (predictors) are listed on the left and represented by the circles which are the input nodes. The middle layer of circles represent the hidden layer with the circles representing the hidden nodes. The far right single circle represents the output node that serves to predict clinical deterioration from the neural network.