Here’s a discouraging factoid for you: Despite projected industry-wide job growth that’s much higher than average (19% from 2016 to 2026), dentistry is expected to face a workforce shortage in certain parts of the United States. At the end of 2017, the Health Resources and Services Administration (HRSA) identified 5,866 dental health professional gap areas — where a population falls below one dentist per 5,000 people — that are home to nearly 63 million people.
Orthodontists are particularly overburdened with paperwork brought on by overfull appointment schedules. But what if AI could lend a hand? Enter research conducted by scientists and graduate students at Osaka University in Japan, who wrote a paper (“Using Natural Language Processing to Develop an Automated Orthodonic Diagnostic System“) proposing a system that uses natural language processing to design an orthodontic treatment plan from findings included in dentist-penned medical certificates. They say its treatment priority rankings match those of human dentists over half the time.
“An [AI] system that can implement the years of experience of a specialist would be of great significance in providing patients with evidence-based medical care,” wrote the coauthors. “[A]utomatic summarization of orthodontic diagnoses or presentation of necessary examinations in an orthodontic clinic would reduce the heavy workload of dentists, as well as help less experienced dentists in avoiding oversights and judgment errors.”
To that end, the researchers’ AI extracts imaging and modeling findings from records that represent a patient’s medical condition, along with diagnoses (listed by priority) and treatment plans relevant to these findings. It organizes each condition mentioned in the treatment protocol, taking into account around 400 types of condition labels, and then performs machine-learned ranking of each extracted medical condition before generating relevant sentence pairs and outputting a list of treatment priorities for each problem.
To train the system, researchers compiled a data set of 990 documents written by dentists, with the conditions and problems in each organized into 423 classes. (About 180 documents were set aside for validation and evaluation.) Each sentence was divided into words (for a vocabulary size of 2,075) for a training corpus of 146 vocabulary and 320 classes, and the overall performance of the AI system was evaluated using an F1-score, or the harmonic average of the precision and the recall.
The team reports that the best-performing model achieved an F1-score of 0.585, which they note leaves room for improvement. In future work, they intend to implement robust sentence encoding for incomplete sentences, findings for treatment prioritization, and text simplification for everything from protocol summaries to consent form documents.