Using Machine Learning Predictive Models to Aid Manufacturers Identify Early Warning Signs of Ventilator Defects

The United States (U.S.) Food and Drug Administration (FDA) maintains a Manufacturer and User Facility Device Experience (MAUDE) database containing end- user experience reports, which detail adverse events encountered with medical devices. However, there is currently no process for efficiently mining through the millions of reports available. Furthermore, the current Coronavirus Disease 2019 pandemic has exacerbated the need for an innovative technique to preemptively detect device malfunctions for essential medical devices, such as ventilators. This need was evidenced by a recent mass recall of faulty ventilators, one of the most crucial devices through the pandemic, costing the manufacturer millions of dollars and debilitating an already-frail supply chain.This praxis explores a combined machine learning approach utilizing Latent Dirichlet Allocation (LDA) topic modeling and predictive models for text classification—including Logistic Regression (LR), Gaussian Naïve Bayes (NB), Random Forest (RF), Adaboost (AB) and Support Vector Machines (SVM) with Stochastic Gradient Descent training (SGD)—to uncover early warning signs of ventilator defects and to predict potential recalls and their associated recall classifications. The proposed methodology presented in this praxis introduces a novel approach for ventilator defect detection and prediction of ventilator recalls. Using LDA topic modeling, early warning signs of ventilator defects are extracted from relevant adverse event narrative reports included in the MAUDE database and then utilized as feature vectors for recall status determination through predictive modeling. Finally, a text classification model is employed to determine the applicable recall classification level of ventilators that have been predicted to be recalled. The results from this research indicate that ventilator manufacturers and design engineers can effectively uncover early warning signs of ventilator defects and proactively develop recall mitigation strategies using the proposed methodology.

Author

• Santiago-Pérez, Valerie

Language

• en

Keyword

• predictive modeling
• machine learning
• LDA topic modeling
• defect discovery

Date created

• 2022

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Engineering Management

Degree

• D.Engr.

Advisor

• Bojanić, Oliver

Committee Member(s)

• Blackford, J.P.
• Etemadi, Amir

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/pk02cb63c