Prediction of multiclass surgical outcomes in glaucoma using multimodal deep learning based on free-text operative notes and structured EHR data
Presenters
Michelle Hribar, Phd
Wei-Chun Lin, MD, PhD
Statement of Purpose
The integration of the vast available EHR data and artificial intelligence (AI) in healthcare management and clinical care has been a growing area in recent decades. In this study, we focused on predicting the outcomes of glaucoma surgery, where long-term outcomes are critical for effective postoperative care. Unlike previous models that primarily focused on binary outcomes using either structured pre-operative or post-operative data, our study introduces a comprehensive approach by predicting multiclass surgical outcomes. These multiclass predictions provide a nuanced understanding of potential postoperative scenarios, enabling tailored patient care.
With the rapid advancement of Natural Language Processing (NLP) development, there has been increasing attention on the utilization of free-text clinical data. We developed multimodal deep learning models that integrate both structured EHR data and the information contained in free-text operative notes. Additionally, we explored different methods for effectively extracting information from free-text notes, such as using a pre-trained Bio-Clinical BERT model and custom word embedding with a transformer block. Overall, our findings demonstrate that the integration of intraoperative details can enhance model performance, highlighting the untapped potential of operative notes in surgical outcome prediction. Furthermore, more nuanced predictions may assist in improving postoperative patient care.
Learning Outcomes
- Compare information extraction techniques using BioClinical BERT and custom word embeddings for specific types of clinical notes
- Explain the potential of using multimodal models in healthcare outcome prediction
- Describe the importance of multiclass surgical outcome prediction in postoperative care