Artificial intelligence in onco-anaesthesia: Current applications, challenges, and future directions

Table 1 Comparative analysis of artificial intelligence algorithms in onco-anaesthesia applications

Algorithm type	Primary mechanism	Key clinical application	Advantages	Limitations	Performance metrics
Random forest	Ensemble learning using multiple decision trees	Preoperative risk stratification; Metastasis prediction (MoLPre)	Handles non-linear data; robust to outliers; explainable	Can overfit; less effective for real-time waveforms	AUROC 0.73-0.77[6], AUC 0.92 (internal validation), 0.90 (external validation)[17]
Gradient boosting	Iterative ensemble correcting prior errors	Mortality prediction; delirium prediction	High performance; handles missing data	Computationally intensive	AUC 0.85[31]; AUC 0.840[11]
Stacking ensemble	Meta-learning combining RF, XGBoost, SVM	Early recurrence (NSCLC)	Improved generalization	Complex implementation	AUC 0.81 (test), 0.80 (external validation)[16]
CNN	Deep learning for grid-like data (images)	Ultrasound guidance; radiomics	State-of-the-art image recognition	Low interpretability; needs large datasets	95%-100% accuracy[24]
RNN/LSTM	Deep learning for sequential data	Hemodynamic prediction (HPI); Nociception	Excellent for waveform analysis	High computational cost	AUROC 0.94, Accuracy 0.88[22]
BiLSTM with Attention	Bidirectional sequence processing	Depth of anesthesia (EEG)	Superior temporal extraction	Computationally intensive	88.7% accuracy[3]
Gaussian process regression	Bayesian non-parametric regression	Remifentanil PK prediction	Uncertainty estimates	Intensive for large datasets	R² = 0.96[3]
Reinforcement learning	Trial-and-error optimization	Closed-loop anesthesia	Adapts to patient variability	Safety concerns during learning	Correlation 0.88[3]
NLP	Human language processing	EHR phenotyping	Unlocks unstructured text	Struggles with medical jargon	Variable[6]

RF: Random forest; AUC: Area under curve; AUROC: Area under receiver operating characteristic curve; NSCLC: Non-small cell lung cancer; SVM: Support vector machine; CNN: Convolutional neural network; RNN: Recurrent neural network; LSTM: Long short-term memory; BiLSTM: Bidirectional long short-term memory; HPI: Hypotension Prediction Index; EEG: Electroencephalogram; PK: Pharmacokinetic; NLP: Natural language processing.

Table 2 Artificial intelligence-driven predictive models for oncological outcomes

Outcome domain	Specific target	AI methodology	Key findings/performance	Clinical implication
Hemodynamics	Intraoperative hypotension	Gradient boosting; CNN-RNN	AUROC 0.94; Accuracy 0.88; 6 endotypes[21,22]	Preemptive endotype-specific treatment
Hemodynamics	Anastomotic leakage (esophagectomy)	Threshold analysis	MAP < 65 mmHg: OR 1.02 per 10-minute[20]	Avoid prolonged hypotension
Complications	Anastomotic leakage (colorectal)	ANN	Superior sensitivity vs surgeons[13]	Early intervention/diversion
Complications	CD ≥ 3 complications	XGBoost with SHAP	AUROC 0.77 (train), 0.73 (test)[6]	Transparent risk stratification
Neurological	Postoperative delirium	XGBoost (AI-delirium guard)	AUC 0.85[31]	Targeted prevention; web tool1
Pain	Chronic postsurgical pain	Stacking ensemble (EHR + Wearable)	AUC 0.90; 89% accuracy[35]	Precision analgesia intervention
Prognosis	Recurrence (NSCLC)	Cox regression	Stage I: P = 0.036; stage II-III: NS[25]	Stage-specific counseling
Prognosis	Metastasis (cT1 lung)	Random forest (MoLPre)	AUC 0.92 (internal validation), 0.90 (external validation)[17]	Online prediction tool2
Prognosis	Early recurrence (NSCLC)	Stacking ensemble	AUC 0.81 (test), 0.80 (external validation)[16]	Adjuvant therapy decisions
Liver surgery	Post-Hepatectomy failure	Multimodal ML (radiomics + clinical)	Outperformed MELD/Child-Pugh[14]	Resection vs transplant selection

¹https://ai-delirium-guard.streamlit.app.

²https://molpre.cqmu.edu.cn/.

AI: Artificial intelligence; CNN: Convolutional neural network; RNN: Recurrent neural network; AUROC: Area under receiver operating characteristic curve; MAP: Mean arterial pressure; OR: Odds ratio; ANN: Artificial neural network; CD: Clavien-Dindo; SHAP: SHapley Additive exPlanations; AUC: Area under curve; NSCLC: Non-small cell lung cancer; NS: Not significant; ML: Machine learning; MELD: Model for end-stage liver disease.