Development of Natural Language Processing Algorithm for Dental Charting

The purpose was to develop a natural language processing algorithm, which will input data from oral examination transcripts to a structured database. Four case vignettes were produced for primary, mixed, and permanent dentition patients with varying degree of soft tissue pathology, caries, existing restorations and occlusion relationships. After calibration, dental students were instructed to perform simulated oral examinations based on the case vignettes, using natural language as they would to an assistant who was charting the case. The investigator transcribed the audio recordings of the simulated oral exams. Twenty simulated oral examinations were collected to develop a natural language processing algorithm in JAVA. The algorithm was reviewed and refined. Four additional simulated oral examinations were performed and transcribed as validation. Volunteers were asked to read the validation transcripts and fill in paper dental charts accordingly. The accuracy of the human volunteers and the algorithm were calculated. The case vignettes had an average of 56 data points. After improvements, the recall rate of algorithm extracting data from transcripts was 99.0% ± 3.3% and the precision was 97.8% ± 4.1%. For the validation transcripts, human subjects had a 100% recall and precision rate. The mean recall and precision of algorithm processing the validation transcripts were 98.4% ± 3.2% and 98.3% ± 1.9%, respectively. There was no statistic difference between human and the algorithm. Natural language processing algorithm performs comparably with humans. The natural language processing algorithm potentially serves as a starting point to implement speech recognition for a voice-activated automatic dental charting system.