What Drives Length of Stay After Elective Spine Surgery? Insights from a Decade of Predictive Modeling

Technical Summary: What Drives Length of Stay After Elective Spine Surgery? Insights from a Decade of Predictive Modeling

Problem Statement
Predicting the Length of Stay (LOS) for patients undergoing elective spine surgery is critical for optimizing patient outcomes, managing hospital capacity, and controlling costs, which average approximately $2,883 per inpatient day. While computational methods like Artificial Intelligence (AI) and Machine Learning (ML) have advanced, it remains unclear which models offer superior precision and accuracy for LOS prediction in this specific domain. Furthermore, existing systematic reviews have either focused on diverse outcomes (not exclusively LOS), lacked detailed comparisons between ML and traditional statistical approaches, or failed to differentiate the specific predictors driving model performance. There is a noted gap in the literature regarding standardized outcome definitions, external validation, and the comparative analysis of model performance specifically for elective spine surgery.

Methodology
This study is a systematic review conducted in accordance with PRISMA guidelines.

• Search Strategy: The authors searched PubMed, Google Scholar, and the ACM Digital Library for studies published between December 1, 2015, and December 1, 2024.
• Inclusion Criteria: Original empirical studies predicting LOS in elective spine surgery patients using data-driven techniques (statistical or ML models), published in English with full-text access.
• Selection Process: From an initial pool of 1,263 studies, 29 met the inclusion criteria after screening titles, abstracts, and full texts.
• Data Extraction: Two independent reviewers extracted data using a comprehensive codebook covering study characteristics, modeling approaches (statistical vs. ML vs. NLP), outcome definitions (continuous, binary, or percentile-based), predictor variables, and performance metrics (AUC, accuracy, precision, recall, F1 score).
• Quality Assessment: The risk of bias was assessed using the Cochrane risk of bias tool.

Key Contributions

• Focused Scope: Unlike prior reviews (e.g., Lubelski et al.) that covered broad outcomes, this study exclusively analyzes LOS prediction in elective spine surgery.
• Temporal Expansion: It includes 16 additional papers published after 2020, capturing recent advancements in AI methodologies and data integration.
• Granular Predictor Analysis: The study systematically categorizes and analyzes a vast array of predictors (demographic, clinical, operational, and contextual) to identify which variables most significantly impact model performance.
• Comparative Performance Analysis: It provides a direct comparison of performance scores across various models, including logistic regression, random forests, boosting algorithms, and neural networks, identifying the best-performing models based on key metrics.

Results

• Study Characteristics: The 29 included studies were predominantly retrospective (90%), with sample sizes ranging from 77 to 453,717 patients. Data sources included Electronic Health Records (44.8%), national databases like NSQIP (20.7%), and state-level databases. Most studies (72.4%) were published between 2020 and 2024.
• Modeling Approaches:
  • Statistical Models: Logistic regression was the most common approach (68.97% of studies), often serving as a baseline.
  • Machine Learning: 24.1% of studies employed ML methods, including Random Forests, Gradient Boosting (XGBoost, LightGBM), Support Vector Machines (SVM), and K-Nearest Neighbors (KNN).
  • Hybrid Approaches: 17.24% of studies compared statistical and ML methods.
• Performance:
  • ML models generally outperformed traditional statistical models, with Area Under the Curve (AUC) values ranging from 0.94 to 0.99 in top-performing studies.
  • Top Performers: K-Nearest Neighbors (KNN) and Naive Bayes with adaptive boosting achieved the highest reported AUCs (0.99 and 0.94, respectively). Random Forest and Gradient Boosting models also demonstrated superior performance, frequently exceeding an AUC of 0.80.
  • Logistic regression models showed respectable but generally lower performance, with AUCs typically between 0.70 and 0.80.
• Key Predictors:
  • Demographics: Age was the most significant predictor (included in 79.3% of studies), with older patients associated with longer LOS.
  • Clinical Factors: Comorbidities, specifically hypertension, diabetes, and cardiovascular conditions, were strong predictors included in 79.3% of studies.
  • Surgical Factors: Operation duration and estimated blood loss were critical, linked to extended LOS in 72.4% and 55.2% of studies, respectively. Minimally invasive procedures were associated with shorter LOS in 62.1% of studies.

Significance and Claims
The paper claims that while ML models offer strong potential for improving discharge planning and hospital resource management, their clinical utility is currently limited by a lack of standardization and external validation.

• Current State: The review highlights that ML models, particularly ensemble methods and specific algorithms like KNN and Naive Bayes, hold the potential to outperform traditional statistical approaches in handling complex, non-linear relationships in spine surgery data.
• Limitations: The authors note that many studies lack robust external validation, use small sample sizes, and suffer from inconsistent definitions of LOS outcomes, making direct comparisons difficult.
• Future Directions: The paper advocates for future studies to prioritize standardized outcome definitions, transparent reporting, and the incorporation of unstructured data (e.g., clinical notes via NLP/LLMs) to capture nuanced patient histories. It suggests that hybrid deep learning models and advanced architectures (e.g., GNNs, attention mechanisms) represent a promising future direction, though these are not yet standard in the reviewed spine surgery literature.

In conclusion, the study positions data-driven predictive modeling as a vital tool for optimizing elective spine surgery care but emphasizes that widespread real-world deployment requires addressing current gaps in validation and data standardization.