Artificial Intelligence-Based Multimodal Risk Assessment Model for Surgical Site Infection (AMRAMS): Development and...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Surgical Outcomes Research

Background:

• Surgical site infections (SSIs) are common healthcare-associated infections, increasing mortality, hospital stays, and costs.
• Existing SSI risk assessment models often lack sufficient accuracy for effective preoperative patient identification and intervention.
• There is a need for improved predictive models to guide clinical decisions and reduce SSI incidence.

Purpose of the Study:

• To develop an Artificial Intelligence-based Multimodal Risk Assessment Model for Surgical site infection (AMRAMS) using routinely collected clinical data.
• To validate the model's predictive discrimination internally and externally.
• To compare AMRAMS performance against the National Nosocomial Infections Surveillance (NNIS) risk index.

Main Methods:

• Utilized electronic medical record data from inpatients between 2014-2019.
• Incorporated patient demographics, preoperative lab results, and free-text clinical notes.
• Employed word-embedding techniques and trained various machine learning models, including CNN and self-attention networks, for risk prediction.

Main Results:

• Convolutional Neural Network (CNN) achieved the highest internal validation AUROC (0.889), outperforming other models.
• Self-attention network showed strong performance in both internal (0.882) and external (0.879) validation.
• AMRAMS models significantly outperformed the surgeon-scored NNIS risk index (AUROC 0.651).

Conclusions:

• The developed AMRAMS, leveraging deep learning (CNN, self-attention) and EMR data, significantly improves SSI risk prediction accuracy.
• Incorporating semantic embeddings from preoperative notes enhances model performance.
• AMRAMS offers a more accurate, personalized alternative to the NNIS index for preoperative SSI intervention guidance.