# Artificial Intelligence Approaches to Predict Postoperative Length of Hospital Stay in Head and Neck Cancer Patients: A Systematic Review 

## TL;DR

This study reviews AI models used to predict hospital stay lengths for head and neck cancer patients, finding moderate accuracy but significant room for improvement.

## Contribution

The paper systematically evaluates AI models for predicting hospital length of stay in head and neck cancer surgery patients, highlighting methodological gaps.

## Key findings

- AI models showed moderate accuracy (AUROC 0.66-0.80) in predicting hospital length of stay.
- Common variables like age, BMI, and comorbidities were used, but not always most predictive.
- All studies had high risk of bias, with few addressing data balancing or external validation.

## Abstract

Background/Objectives: The aim of the present systematic review is to evaluate the performance of AI models for length of stay prediction. Methods: This SR was carried out in accordance with PRISMA 2020 and registered in PROSPERO database (CRD420251039985). Using the PICOS framework, we formulated the following research question: “Can artificial intelligence models accurately predict hospital length of stay (LOS) in patients undergoing head and neck (H&N) cancer surgery?” We searched the Cochrane Library, Embase, PubMed, and Scopus, with additional gray literature identified through Google Scholar and ProQuest. Risk of bias (RoB) was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST), and a narrative synthesis was performed to summarize qualitative findings. Results: Of 1304 identified articles, 5 met inclusion criteria, covering 5009 patients. All studies used supervised learning to predict LOS with different variables presenting stronger associations with increased hospital LOS. Age, race, ASA score, BMI, and comorbid factors like smoking and arterial hypertension were comon variables across studies but not always the ones most strongly associated with LOS. One study also predicted discharge to non-home facilities and prolonged LOS; only one applied data balancing. Model accuracies ranged from 0.63 to 0.84, and area under the receiver operator characteristics curve (AUROC) values from 0.66 to 0.80, suggesting moderate discriminative performance. All studies had a high risk of bias, though no applicability concerns were noted. Conclusions: AI models show potential for LOS prediction after H&N cancer surgery; however, an elevated RoB and methodological shortcomings constrain the current evidence. Methodological improvements, external validation, and transparent reporting is essential to enhance reliability and generalizability, enabling integration into clinical decision-making.