Use of Deep Learning to Identify Peripheral Arterial Disease Cases From Narrative Clinical Notes.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Health

Background:

• Peripheral arterial disease (PAD) is a leading cause of amputation with low patient and provider awareness.
• Current methods for identifying PAD patients, such as keyword search (KWS), have limitations in flexibility and accuracy.
• There is a need for advanced strategies to effectively screen and identify individuals with PAD.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL) models in identifying patients with peripheral arterial disease (PAD) from unstructured electronic health record (EHR) data.
• To compare the performance of DL models against traditional keyword search (KWS) algorithms for PAD case detection.

Main Methods:

• Utilized EHR data from a statewide health information exchange to create a PAD patient dataset.
• Developed a DL model using a BioMed-RoBERTa base, fine-tuned on a training cohort (70% of data).
• Compared the DL model's performance against a state-of-the-art KWS algorithm on a held-out testing cohort (30% of data).

Main Results:

• The DL model demonstrated superior performance over KWS across all measured metrics, including sensitivity, specificity, and accuracy.
• DL achieved a sensitivity of 0.70 versus 0.62 for KWS, and an accuracy of 0.96 versus 0.91, respectively (P < 0.001).
• The DL model's specificity (0.99) and negative predictive value (0.97) also significantly surpassed KWS.

Conclusions:

• Deep learning models offer a more effective approach for identifying peripheral arterial disease (PAD) cases from clinical narratives compared to keyword search.
• This study highlights the potential of advanced NLP techniques for improving PAD patient identification within EHR systems.
• Future research will focus on developing DL models for staging PAD severity using clinical scoring systems.