A cost-sensitive multiclass machine learning framework for postoperative neurosurgical triage (Neuro-TACTIC)

Area of Science:

• Neurosurgery
• Machine Learning
• Healthcare Resource Management

Background:

• Postoperative patient placement is crucial for balancing safety and resources.
• Current models often use a binary ICU vs. non-ICU decision, lacking adaptability.
• Existing methods fail to account for local resource constraints or specific definitions of critical care events.

Purpose of the Study:

• To develop a cost-sensitive machine learning framework (Neuro-TACTIC) for stratifying neurosurgical patients into three postoperative monitoring levels.
• To create a model that can adjust risk thresholds based on local resource availability and definitions of critical care.
• To move beyond binary ICU/non-ICU decisions for more precise patient care allocation.

Main Methods:

• Developed Neuro-TACTIC, an XGBoost-based classifier utilizing 27 features (demographic, intraoperative, imaging).
• Trained on a retrospective cohort of 1072 patients undergoing elective craniotomy.
• Incorporated a tunable parameter (ζ) to balance resource costs and harm costs, enabling adjustment of over- and under-triage.

Main Results:

• The framework demonstrated stable performance across various cost settings in cross-validation and bootstrap analyses.
• At ζ=0.975, performance metrics included AUCμ=0.67±0.03 and weighted F1=0.49±0.03 in the development cohort.
• Independent validation showed AUCμ=0.60±0.04 and weighted F1=0.44±0.06, with operative duration, tumor volume, and surgical position as key predictors.

Conclusions:

• Neuro-TACTIC shows the feasibility of cost-sensitive, three-tier postoperative triage modeling in neurosurgery.
• The model offers a more nuanced approach to postoperative care allocation than traditional binary methods.
• Prospective validation and multicenter evaluation are necessary before clinical implementation.