DEVELOPMENT OF MACHINE LEARNING MODELS TO PREDICT RISK OF PATHOLOGY OR NEED FOR INTERVENTION AMONGST ADULT PATIENTS UNDERGOING COLONOSCOPY

Aims Background The number of colonoscopies performed yearly is constantly increasing around Europe. Accordingly, endoscopy services are coming under huge service-delivery pressure, and waiting times are becoming unacceptable; This calls for better risk-stratification.

Aim To construct foundational machine-learning models which predict the likelihood of pathology and need for intervention at colonoscopy.

Methods Colonoscopy records were anonymised. Predictors available within the dataset included: sex, age, procedure indication and urgency. Outcomes of interest included: ´all-pathology´ and ´interventional-pathology´, where intervention (including biopsy) was required. Odds Ratios/Chi-Square statistics were calculated for all predictors. We then developed and internally validated multivariate logistic regression (LR) models, decision tree classifiers and artificial neural nets.

Results 23,663 colonoscopies, performed on 18,677 individual patients were analysed. Mean age: 60.89; 50.91% Female. The largest procedure indications within the cohort were: Polyp Surveillance[n=6137] and Bowel Cancer Screening (BCSP)[n=4508]. 74.92%(+/-0.56%-95%CI) of tests contained pathology; 64.84%(+/-0.61%-95%CI) contained ‘interventional-pathology’. Some key predictors all with p< 0.01 significance are listed in [[Table 1]]:

Tab. 1 Predictor Statistics

Predictor	´All-Pathology´ OR(95%CI)	´All-Pathology´ Pearson-Chi2	´Interventional-Pathology´ OR(95%CI)	´Interventional-Pathology´ Pearson-Chi2
Age: >55	3.51(3.30-3.74)	1661.67	1.09(1.02-1.15)	7.49
Sex: Male	1.77(1.67-1.88)	2004.02	0.71(0.67-0.74)	214.07
Indication: Previous-Polyps	3.34(3.07-3.64)	836.47	2.42(2.26-2.59)	672.50
Indication: BCSP	3.94(3.55-4.38)	738.61	1.33(1.23-1.42)	63.05

Best case scenario for finding pathology: 97.6%(+/-0.93%-95%CI) of males, Aged>55, undergoing urgent colonoscopy for polyp surveillance, while only 21.79%(+/-6.06%-95%CI) of females, < 55yrs, undergoing urgent colonoscopy for anaemia had any ‘interventional-pathology’. Given the nature of the dataset, LR performed best and was able to obtain an optimised AUC of 0.75 for predicting ‘all-pathology’ and 0.70 for ‘interventional-pathology’.

Conclusions Machine-learning algorithms can use simple pre-procedure parameters to predict the likelihood of finding significant pathology during colonoscopy; This can help us better utilise colonoscopy resources, potentially reducing/replacing 25% of current colonoscopy workload, with significant resultant health-economic benefits.