Developing Children’s Oral Health Assessment Toolkits Using Machine Learning Algorithm

Abstract

Objectives:

Evaluating children’s oral health status and treatment needs is challenging. We aim to build oral health assessment toolkits to predict Children’s Oral Health Status Index (COHSI) score and referral for treatment needs (RFTN) of oral health. Parent and Child toolkits consist of short-form survey items (12 for children and 8 for parents) with and without children’s demographic information (7 questions) to predict the child’s oral health status and need for treatment.

Methods:

Data were collected from 12 dental practices in Los Angeles County from 2015 to 2016. We predicted COHSI score and RFTN using random Bootstrap samples with manually introduced Gaussian noise together with machine learning algorithms, such as Extreme Gradient Boosting and Naive Bayesian algorithms (using R). The toolkits predicted the probability of treatment needs and the COHSI score with percentile (ranking). The performance of the toolkits was evaluated internally and externally by residual mean square error (RMSE), correlation, sensitivity and specificity.

Results:

The toolkits were developed based on survey responses from 545 families with children aged 2 to 17 y. The sensitivity and specificity for predicting RFTN were 93% and 49% respectively with the external data. The correlation(s) between predicted and clinically determined COHSI was 0.88 (and 0.91 for its percentile). The RMSEs of the COHSI toolkit were 4.2 for COHSI (and 1.3 for its percentile).

Conclusions:

Survey responses from children and their parents/guardians are predictive for clinical outcomes. The toolkits can be used by oral health programs at baseline among school populations. The toolkits can also be used to quantify differences between pre- and post-dental care program implementation. The toolkits’ predicted oral health scores can be used to stratify samples in oral health research.

Knowledge Transfer Statement:

This study creates the oral health toolkits that combine self- and proxy- reported short forms with children’s demographic characteristics to predict children’s oral health and treatment needs using Machine Learning algorithms. The toolkits can be used by oral health programs at baseline among school populations to quantify differences between pre and post dental care program implementation. The toolkits can also be used to stratify samples according to the treatment needs and oral health status.