An explainable long short-term memory network for surgical site infection identification

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Surgical outcomes research

Background:

• Current surgical site infection (SSI) surveillance relies on manual chart review, which is time-consuming.
• Machine learning (ML) offers automated identification of SSIs from electronic health records.
• Deep learning (DL) models, while powerful, often lack interpretability.

Purpose of the Study:

• To develop and evaluate an explainable deep learning model for SSI identification.
• To compare the performance of the DL model against traditional ML methods.
• To enhance the interpretability of DL models in healthcare surveillance.

Main Methods:

• Utilized structured data and clinical notes from electronic health records (EHRs) from January 2016 to June 2021.
• Developed a long short-term memory (LSTM) network with an attention layer for SSI detection.
• Compared LSTM performance against traditional ML models using metrics like area under the receiver operating characteristic curve (AUROC) and sensitivity.

Main Results:

• SSIs were present in 4.7% of 9,185 operative events.
• The LSTM model achieved higher AUROC (0.954) and sensitivity (0.920) than the top traditional model (AUROC: 0.937, sensitivity: 0.736).
• Key features identified by the LSTM model included procedure codes and laboratory values.

Conclusions:

• Explainable deep learning, specifically the LSTM model, demonstrates superior performance for SSI surveillance compared to traditional ML.
• The developed model offers a potential replacement for manual chart review, increasing efficiency.
• Automated, explainable SSI surveillance can significantly contribute to reducing infection rates.