IADR Abstract Archives
Machine Learning Approach in Predicting Restoration Failure Utilizing Electronic Health Records
Objectives: This study aimed to apply a Machine Learning approach within Electronic Health Records (EHRs) to identify predictive factors for restoration failure in dental practice.
Methods: Secondary data from the BigMouth Repository, a centralized oral health data repository utilizing electronic health records from nine dental schools across the U.S., were analyzed. The study included data from permanent molars restored with composite or amalgam. Restoration failure was determined based on whether the restored tooth necessitated Root Canal Treatment or extraction within five-year.
We employed a neural network Machine Learning (ML) approach using input factors available in the BigMouth dataset, which included the type of restoration, caries depth, number of surfaces involved, tooth number, follow-up time, age, and gender. Other factors that contribute to restoration failure, such as pre-diagnosis, usage of liner, isolation, and the type of composite materials, were not present in the BigMouth repository.
We conducted three separate ML modelling: one combining both amalgam and composite, one exclusively for amalgam, and the last for composite alone. Our dataset was divided into 80% for training the ML model and 20% for testing.
Results: Our ML analysis achieved 70% accuracy for predicting the risk of restoration failure. Our analysis highlighted that amalgam restoration was more susceptible to failure than composite restoration. Feature importance analysis revealed that the most influential parameter to amalgam failure was tooth number, whereas caries depth played a more crucial role in the composite failure.
Conclusions: This secondary analysis suggests a superior survival rate for composite restoration compared to amalgam restoration. However, it's crucial to acknowledge that significant variables related to restoration failure were missing from this dataset. To enhance AI and prediction capabilities in dentistry, academic institutions should reinforce the implementation of diagnostic codes and improve their reporting for more precise AI application.
Methods: Secondary data from the BigMouth Repository, a centralized oral health data repository utilizing electronic health records from nine dental schools across the U.S., were analyzed. The study included data from permanent molars restored with composite or amalgam. Restoration failure was determined based on whether the restored tooth necessitated Root Canal Treatment or extraction within five-year.
We employed a neural network Machine Learning (ML) approach using input factors available in the BigMouth dataset, which included the type of restoration, caries depth, number of surfaces involved, tooth number, follow-up time, age, and gender. Other factors that contribute to restoration failure, such as pre-diagnosis, usage of liner, isolation, and the type of composite materials, were not present in the BigMouth repository.
We conducted three separate ML modelling: one combining both amalgam and composite, one exclusively for amalgam, and the last for composite alone. Our dataset was divided into 80% for training the ML model and 20% for testing.
Results: Our ML analysis achieved 70% accuracy for predicting the risk of restoration failure. Our analysis highlighted that amalgam restoration was more susceptible to failure than composite restoration. Feature importance analysis revealed that the most influential parameter to amalgam failure was tooth number, whereas caries depth played a more crucial role in the composite failure.
Conclusions: This secondary analysis suggests a superior survival rate for composite restoration compared to amalgam restoration. However, it's crucial to acknowledge that significant variables related to restoration failure were missing from this dataset. To enhance AI and prediction capabilities in dentistry, academic institutions should reinforce the implementation of diagnostic codes and improve their reporting for more precise AI application.