EP817 - Predicting Optimal Length of Stay After Colectomy Using Machine Learning Models

Monday, May 12, 2025

1:36 PM – 1:41 PM PT

ePoster Presenter(s)

AA

Amir Ashraf ganjouei, MD, MPH

Research Fellow
University of California, San Francisco
San Francisco, CA, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.

Author(s)

SH

Sophia Hernandez, MD

Resident
University of California, San Francisco
San Francisco, California, United States

Disclosure information not submitted.
FR

Fernanda Romero-Hernandez, MD

Resident
University of California, San Francisco
San Francisco, California, United States

Disclosure information not submitted.
Elizabeth C. Wick, MD

Professor of Surgery
University of California, San Francisco
University of California, San Francisco
San Francisco, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
EF

Emily VA Finlayson, MD, MS

Professor of Surgery
University of California, San Francisco
San Francisco, California, United States

Disclosure information not submitted.
HC

Hueylan Chern, MD

Professor of Surgery
University of California, San Francisco
San Francisco, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
BZ

B. Mark Zhao, University of California, San Francisco

Assistant Professor of Clinical Surgery
University of California, San Francisco
San Francisco, CA, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.

Senior Author(s)

Madhulika G. Varma, MD

Professor of Surgery
University of California, San Francisco
San Francisco, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no relevant relationships with ineligible companies.

All of the relevant financial relationships listed below have been mitigated.

Amir Ashraf ganjouei, MD, MPH: This individual has no financial relationships with ineligible companies.

Elizabeth C. Wick, MD: This individual has no financial relationships with ineligible companies.

Hueylan Chern, MD: This individual has no financial relationships with ineligible companies.

B. Mark Zhao, University of California, San Francisco: This individual has no financial relationships with ineligible companies.

Madhulika G. Varma, MD: This individual has no relevant relationships with ineligible companies.

Image or Table (Do Not add text here):

Purpose/Background: As healthcare systems contend with increasing demands for inpatient resources, optimizing hospital length of stay (LOS) after elective surgeries has become critical to increase bed availability for high-acuity patients. The introduction of enhanced recovery protocols have significantly reduced LOS after colectomy in the past decade. However, there remains no widely available tool to predict the likelihood of achieving the optimal LOS following elective colectomy. Such predictions are essential for efficient resource utilization and cost management, as well as for improving patient care through tailored postoperative planning. Therefore, our aim was to develop and validate a machine learning-based tool to predict optimal LOS after colectomy.

Methods/Interventions: Adult patients who underwent colectomy between 2014 and 2021 were identified from the ACS-NSQIP dataset. Patients were excluded if they underwent emergent surgery. Optimal LOS was defined as discharge within three days postoperatively. Three machine learning models (neural network, random forest, and XGBoost) were trained and compared to a traditional logistic regression model. Each model was trained on 80% of the dataset and validated on the remaining 20%. The recursive feature elimination algorithm was employed to optimize the number of input variables, reduce overfitting, and enhance interpretability. Finally, SHAP (SHaply Additive explanation) values were calculated to determine the importance of each variable in predicting optimal LOS.

Results/Outcomes: Among the 184,286 colectomy cases analyzed, the mean age was 61 years, and 51% were female. Of these patients, 31,015 (16.8%) had an optimal LOS. After training the models using 80% of the data, the XGBoost model achieved the highest area under the ROC curve (AUC) in the test dataset, with an AUC of 0.76 (CI: 0.75-0.77), followed by the neural network (AUC 0.75, CI: 0.74-0.76), random forest (AUC 0.73, CI: 0.72-0.74), and logistic regression (AUC 0.71, CI: 0.70-0.72, p-value < 0.001 for comparison against XGBoost) (Figure 1A). SHAP value analysis identified the most important predictors of an optimal LOS, which included undergoing minimally invasive surgery, receiving preoperative antibiotic bowel preparation, higher preoperative albumin and hematocrit levels, and younger age (Figure 1B).

Conclusion/Discussion: Utilizing advanced machine learning techniques, we developed a robust model that accurately predicts the achievement of an optimal LOS after colectomy. This model identifies critical preoperative factors, such as preoperative bowel preparation, management of malnutrition and anemia, as keys to achieving optimal LOS in elective colectomy cases. This tool may offer clinicians a reliable method to predict postoperative outcomes, facilitate tailored discussions with patients about expected recovery times, and thereby support personalized perioperative care plans and enhance overall recovery strategies. Additionally, by integrating our model into clinical workflows, healthcare providers can improve resource allocation and patient care efficiency.