Artificial Intelligence and Machine Learning in Pediatric Anesthesia: A Narrative Review

Authors
Patrick Fakhoury MS4¹, Aditya Shah MS3¹, Caleb Zimmerman BS¹, Aiman Almasnaah MS4¹, Kush Patel MS4¹, Deepti Sanku MS4¹, Kabir Mendiratta MD⁴, Wael Saasouh MD²˒³

¹ Central Michigan University College of Medicine, Saginaw, MI, USA
² Wayne State University School of Medicine, Detroit, MI, USA
³ Outcomes Research Consortium, Houston, TX, USA
⁴ Detroit Medical Center, Detroit, MI, USA

---

Background

• Pediatric anesthesia is challenging due to small airways, rapid desaturation, limited physiologic reserve, and difficulty assessing pain and distress.

• Traditional tools (age/weight formulas, simple vital-sign thresholds) can be imprecise.

• AI and ML can integrate complex perioperative data to support individualized decision-making.

Objective: To synthesize current applications of AI/ML in pediatric anesthesia and describe their potential to augment perioperative care.

---

Methods

• Narrative review of AI/ML in pediatric anesthesia using CINAHL, PubMed, Scopus, and Web of Science.

• Included original studies in patients 0–18 years involving anesthesia or perioperative care and reporting clinical or predictive outcomes.

• Excluded adult-only studies, mixed cohorts without pediatric data, and purely theoretical or simulation-only work.

• From 893 records screened, 11 studies met inclusion criteria.

---

Results

Key Domains of Application

• Airway Management

  • ML models improved prediction of ETT size and depth vs conventional formulas.

  • Signal-based models accurately classified airway devices (mask, LMA, ETT) and detected intubation/malposition earlier.

• Intraoperative Hypoxemia

  • Gradient boosting and deep learning using high-frequency monitor data predicted hypoxemia (e.g., SpO₂ < 95% within 1 min) more accurately than clinician judgment or fixed thresholds.

• Noninvasive Ventilation Failure

  • Time-aware neural networks predicted BiPAP failure and need for intubation at 6–24 hours, outperforming logistic regression.

• Postoperative Pain and Emergence Delirium

  • AI models reproduced or improved upon traditional pediatric pain scores when predicting need for rescue analgesia.

  • ML-based risk scores predicted emergence delirium with acceptable discrimination and highlighted modifiable perioperative factors.

• Perioperative Risk Stratification

  • Gradient boosting models using demographic, ASA, and procedural variables identified low-risk pediatric patients with high negative predictive value for serious events.

---

Discussion

• Across airway management, hypoxemia prediction, ventilatory support, pain assessment, delirium, and risk stratification, AI/ML generally outperformed rule-based methods and age-based formulas.

• These tools function best as decision-support, not replacements for anesthesiologists, providing earlier warnings and more precise, patient-specific estimates.

• Current evidence is limited by small sample sizes, single-center retrospective designs, and lack of robust external validation.

---

Conclusion

• AI and ML show significant promise in enhancing safety and individualizing care in pediatric anesthesia.

• Greatest benefits to date are seen in airway management and short-term physiologic risk prediction.

• Future work should focus on multicenter prospective validation, integration into OR/EHR workflows, and evaluation of real-world impact on outcomes and clinician workload.