🚨 AI and Neurosurgery 🚨

Introduction:

Following a surgical intervention, the Intensive Care Unit (ICU) admission of patient is critical yet unplanned ICU admission post neurosurgical interventions are often overlooked. The unanticipated ICU admissions and transfers exacerbated the morbidity, increase the length of hospital stay as well as increases the cost of treatment. No standardized, evidence-based protocols exist for post-op ICU triage in neurosurgery, leaving practices reliant on clinician expertise, resulting in variability and missed early interventions. Studies show up to 36% of high-risk ICU cases are missed by human experts and unplanned ICU admission rates in neurosurgery remain between 14% and 28%.

Methods:

• Dataset: Retrospective study of 2,268 neurosurgical patients at University College London Hospital (UCLH).

• NLP Approach: CogStack-MedCAT platform extracts SNOMED-coded clinical concepts from free-text clinical notes (30 days pre-op).

• Machine Learning Models:

  • Random Forest and LSTM with BERT embedding for classification (ward vs. ICU).

  • Explainability: Model interpretability ensured using SHAP and LIME, validated by clinical experts.

Key Results & Performance

• Recall: 0.87 for ICU admissions; 1.0 precision for unplanned ICU admissions.

• AUC (Area Under the Curve): 0.99.

• Over 8,000 clinical concepts extracted, with top predictors being neurovascular, neuro-oncological, and musculoskeletal features.

• Fairness: No significant bias, achieving near-parity across sex and ethnicity groups (0.99 ratio).

Impact:

• Missed ICU cases dropped from 36% (human experts) to just 4% (AI model).

• Improved accuracy in predicting unplanned ICU admissions, enhancing patient safety and resource management.

• Predictive Features: Neurovascular conditions, aneurysms, meningiomas, and osteoporosis emerged as key predictors.

Clinical Relevance & Future Directions:

AI successfully predicts unplanned ICU admissions with high accuracy using routine clinical data.

• Clinical Benefits: Reduced ICU admissions, optimized resource use, lower mortality, and shorter hospital stays.

• Next Steps:

  • Integration into EHR for real-time clinical decision support.

  • Multi-center validation and real-world implementation to further refine and expand the model.

Conclusion: A Paradigm Shift in Neurosurgical Care

This AI-enhanced methodology signifies a transformative leap in personalized medicine, enabling precise identification of neurosurgical patients at heightened risk for unanticipated ICU admission. Leveraging advanced machine learning and natural language processing, the model augments clinical decision-making, streamlines resource allocation & optimizes patient prognostication. With further integration into clinical workflows, AI stands poised to revolutionize neurosurgical care, elevating patient safety and institutional efficiency to unprecedented levels.