Using Natural Language Processing to Develop an Automated Orthodontic Diagnostic System

Tomoyuki Kajiwara, Chihiro Tanikawa, Yuujin Shimizu, Chenhui Chu, Takashi Yamashiro, Hajime NagaharaarXiv 2019 · arXiv

Introduction

Orthodontic diagnosis is a sophisticated process that requires specialists to synthesize diverse diagnostic materials—including X-rays, facial photographs, and dental models—to identify clinical issues and establish a prioritized treatment plan. This process is inherently experience-dependent and prone to subjectivity, particularly for less experienced practitioners. Furthermore, clinical findings are often recorded as unstructured, free-form text, making them difficult for traditional computational tools to analyze.

To address these challenges, researchers from Osaka University have developed an automated orthodontic diagnostic system. This AI-driven framework leverages Natural Language Processing (NLP) to transform clinical notes into structured, prioritized treatment lists, aiming to reduce the cognitive load on dentists and minimize diagnostic errors.

System Architecture

The proposed system follows a multi-stage pipeline that transitions from raw diagnostic input to a structured treatment plan.

Overview of the automated orthodontic diagnostic system pipeline, integrating multi-modal data and NLP for treatment planning.

As illustrated in Figure 1, the system processes patient attribution data alongside visual diagnostic materials. Through automatic recognition and NLP, it extracts specific clinical problems (e.g., "Skeletal 3 due to mandibular retrusion") and subsequently ranks them to assist in treatment planning.

Methodology

The researchers decomposed the diagnostic task into two primary subtasks:

1. Automated Problem Extraction

The system treats clinical notes as input for a multi-label classification task, mapping free-form text to 423 predefined orthodontic problem classes. Two vectorization techniques were evaluated:

  • Bag-of-Words (BoW): A frequency-based approach counting word occurrences.
  • Universal Sentence Encoder (USE): A deep learning model providing dense semantic embeddings. These vectors served as input for a Multi-Layer Perceptron (MLP) to predict the relevant clinical labels.

2. Treatment Prioritization

Once problems are identified, they must be ranked by clinical priority. The team utilized SVM-rank, a supervised learning-to-rank algorithm. They compared three representation methods for the identified problems: BoW, One-of-K (OoK) encoding of class labels, and USE embeddings.

Experimental Results

The study utilized a dataset of 990 medical certificates from the Osaka University Graduate School of Dentistry.

Extraction Performance

In the problem extraction task, the simpler Bag-of-Words model surprisingly outperformed the more complex Universal Sentence Encoder.

Comparison of Precision, Recall, and F1-score for BoW and USE models in the problem extraction task.

As shown in Figure 5, the BoW-based model achieved an F1-score of 0.585, compared to 0.536 for USE. The authors attribute this to the fragmented, bulleted nature of clinical notes, which can be challenging for models trained on standard grammatical prose but are well-handled by keyword-centric approaches like BoW.

Prioritization Performance

In contrast to the extraction results, the ranking task benefited from semantic depth. The USE-based ranking model achieved a Spearman’s rank correlation coefficient ($\rho$) of 0.584, significantly higher than the BoW approach (0.513) and the label-based OoK approach (0.566). This suggests that while identifying a problem is a keyword-matching exercise, determining clinical priority requires a deeper semantic understanding of the conditions.

Conclusion and Future Directions

This research demonstrates the feasibility of using NLP to structure and prioritize orthodontic diagnostic data. While the current system shows promising alignment with human experts, several avenues for future work remain. These include expanding the dataset beyond the initial 990 documents, improving robustness to medical shorthand, and integrating multi-modal data (images and text) for a more holistic diagnostic tool. Additionally, the researchers aim to develop "Automated Simplification" features to translate complex dental terminology into accessible language for patient communication.

Summary updated

BibTeX
@misc{kajiwara2019using,
  author = {Tomoyuki Kajiwara and Chihiro Tanikawa and Yuujin Shimizu and Chenhui Chu and Takashi Yamashiro and Hajime Nagahara},
  title = {{Using Natural Language Processing to Develop an Automated Orthodontic Diagnostic System}},
  year = {2019},
  eprint = {1905.13601},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.13601},
}
Figure