Excellent evidences [[1], [2], [3]] have been produced from the national registry databases, such as the American Joint Replacement Registry (AJRR) [4] in the United States, National Joint Registry (NJR) [5] in the United Kingdom, and Australian Orthopaedic Association National Joint Replacement Registry (AOANJRR) [6]. The Japanese Orthopaedic Association National Registry (JOANR), a large-scale registry system for musculoskeletal diseases, was recently launched to build an evidence-based surgical treatment data. The JOANR is expected to improve medical care quality and optimize medical economy.

Although the JOANR is an important project, it also requires significant effort. Surgeons must register 11 essential data elements on almost all orthopedic procedures and for total hip arthroplasty (THA), in particular, surgeons must register ten additional detailed features. These tasks are based on dedicated sacrifices of surgeons and are naturally free of charge. In addition to maintaining operative records as part of their regular duties, surgeons must record the registry, which requires double effort. One possible solution is to use a system that automatically extracts information about surgeries, which can reduce the burden of registration without the cost of hiring an assistant. Operative records are one of the best targets for extracting data about JOANR because they contain condensed information about surgeries. However, because operative records are written in free text, it has been difficult to handcraft an automated algorithm. The development of natural language processing (NLP), a branch of artificial intelligence, has made it possible to extract features from surgical records. The NLP methods applied in previous studies [[7], [8], [9]] can be classified into rule-based methods, machine learning (ML), which are traditional NLP methods. Wyles et al. [10] have conducted studies that use a rule-based algorithm to extract information on surgical approach and fixation technique from operative records of THA.

In 2018, Google invented the bidirectional encoder representations from transformers (BERT) as the state-of-the-art NLP model [11]. Since ML models consider only the number of occurrences of a word, they cannot recognize relationships between individual words. On the other hand, BERT has a very deep model architecture, based on self-attention layers, allowing the model to learn relationships between individual words. Several studies [12,13] showed that BERT had higher performance than traditional NLP models, and BERT has begun to be implemented in radiology reports. However, to the best of our knowledge, no study has examined benefits of BERT in orthopaedic operative records.

This study aimed to search the best NLP method among a rule-based method, ML, and BERT for building a system that automatically detects some elements in JOANR from THA operative records.