Construction and validation of a machine learning algorithm-based predictive model for difficult colonoscopy insertion

Abstract

Background: Difficulty of colonoscopy insertion (DCI) significantly affects colonoscopy effectiveness and serves as a key quality indicator. Predicting and evaluating DCI risk preoperatively is crucial for optimizing intraoperative strategies.

Aim: To evaluate the predictive performance of machine learning (ML) algorithms for DCI by comparing three modeling approaches, identify factors influencing DCI, and develop a preoperative prediction model using ML algorithms to enhance colonoscopy quality and efficiency.

Methods: This cross-sectional study enrolled 712 patients who underwent colonoscopy at a tertiary hospital between June 2020 and May 2021. Demographic data, past medical history, medication use, and psychological status were collected. The endoscopist assessed DCI using the visual analogue scale. After univariate screening, predictive models were developed using multivariable logistic regression, least absolute shrinkage and selection operator (LASSO) regression, and random forest (RF) algorithms. Model performance was evaluated based on discrimination, calibration, and decision curve analysis (DCA), and results were visualized using nomograms.

Results: A total of 712 patients (53.8% male; mean age 54.5 years ± 12.9 years) were included. Logistic regression analysis identified constipation [odds ratio (OR) = 2.254, 95% confidence interval (CI): 1.289-3.931], abdominal circumference (AC) (77.5-91.9 cm, OR = 1.895, 95%CI: 1.065-3.350; AC ≥ 92 cm, OR = 1.271, 95%CI: 0.730-2.188), and anxiety (OR = 1.071, 95%CI: 1.044-1.100) as predictive factors for DCI, validated by LASSO and RF methods. Model performance revealed training/validation sensitivities of 0.826/0.925, 0.924/0.868, and 1.000/0.981; specificities of 0.602/0.511, 0.510/0.562, and 0.977/0.526; and corresponding area under the receiver operating characteristic curves (AUCs) of 0.780 (0.737-0.823)/0.726 (0.654-0.799), 0.754 (0.710-0.798)/0.723 (0.656-0.791), and 1.000 (1.000-1.000)/0.754 (0.688-0.820), respectively. DCA indicated optimal net benefit within probability thresholds of 0-0.9 and 0.05-0.37. The RF model demonstrated superior diagnostic accuracy, reflected by perfect training sensitivity (1.000) and highest validation AUC (0.754), outperforming other methods in clinical applicability.

Conclusion: The RF-based model exhibited superior predictive accuracy for DCI compared to multivariable logistic and LASSO regression models. This approach supports individualized preoperative optimization, enhancing colonoscopy quality through targeted risk stratification.

Keywords: Colonoscopy; Difficulty of colonoscopy insertion; Least absolute shrinkage and selection operator regression; Logistic regression; Machine learning algorithms; Predictive model; Random forest.

Figure 1

Least absolute shrinkage and selection operator regression variable screening diagram.

Figure 2

Random forest model.

Figure 3

Mean decrease Gini diagram of random forest. BMI: Body mass index; RF: Random forest; SAS: Self-Rating Anxiety Scale.

Figure 4

Receiver operating characteristic curves in training and validation set. A: Receiver operating characteristic (ROC) curves of three models in training set; B: ROC curves of three models in validation set. AUC: Area under the receiver operating characteristic curve; LASSO: Least absolute shrinkage and selection operator; RF: Random forest.

Figure 5

Decision curve analysis curves in training and validation sets. A: Decision curve analysis (DCA) curves of three models in training set; B: DCA curves of three models in validation set. LASSO: Least absolute shrinkage and selection operator; RF: Random forest.

Figure 6

Nomogram prediction model incorporating selected risk factors. SAS: Self-Rating Anxiety Scale.