D1.265 - Predicting Milk Ladder Outcomes for Management of IgE-Mediated Cow’s Milk Protein Allergy: Feature Selection Using Machine Learning Methods

The ‘Milk Ladders’, are forms of dietary advancement therapy (DAT) and are widely used tools designed to induce tolerance in children with or at risk of cow's milk protein allergy (CMPA). 

While studies in Ireland and the UK have demonstrated effectiveness in reintroduction of milk to the diet using DAT-Ladders, there remains wide speculation regarding the safety of the ladders, and whether they are exclusively effective in less sensitised patients who are more likely to naturally outgrow their allergy. Similarly, it has been recommended that DAT-Ladders should be avoided in highly sensitised children or children with a history of anaphylaxis.

While milk ladders have been shown to be safe and effective in retrospective studies, useful clinical predictors of a negative outcome have not been elucidated. Therefore, the aim of this study is to determine predictors of outcomes of DAT using the milk ladders using Machine learning (MachL) in a cohort of Irish children. 

This is a proof-of-concept study applying MachL to retrospective data recollected from patients with IgE-mediated CMPA ML who underwent DAT using the milk ladder (ML). Charts of patients diagnosed with IgE-­mediated CMPA between 2011 and 2020 in the paediatric allergy centre at Cork University Hospital were included. Failure to complete the ladder was defined as the failure to introduce liquid cow’s milk after 36 months of follow-up. Training and test sets were created, and due to class imbalance between those who passed and failed the ladder, SMOTE was applied to the training set before models were applied. Supervised machine learning models were generated using logistic regression (LR) and Random Forest to determine the most important factors for predicting failure of ML completion. 

Following data processing, 171 were included in the final analysis, 142 (83%) successfully completed the ML, while 29 (17%) participants failed to complete the ML. The five most important features for predicting ML outcome with the largest decrease in Gini were the age at the start of treatment, specific IgE to milk, age at diagnosis of CMPA, the duration of breastfeeding, and SPT to milk, while the five least important features with the lowest decrease in Gini were history of concomitant food allergies, family history of any atopic condition, concomitant history of asthma or AD and anaphylaxis to milk before diagnosis. The RF model had an area under the curve (AUC) of 0.855, sensitivity 71.4% and specificity 85.7%, compared to the AUC of the LR model (0.74). 

This study presents the first attempt to apply MachL to predict outcomes of DAT using the milk ladder. The strongest predictors of successful ladder filaure were age at the start of therapy and age at diagnosis, with additional contributions from allergen-specific IgE levels and skin prick test (SPT) results. Feeding practices also appeared relevant, with longer duration of breastfeeding associated with milk ladder outcomes. While rigorous external validation is needed before such algorithms can be integrated into practice, it was demonstrated that age is the most important feature to predict milk outcome and that a history of anaphylaxis prior to diagnosis is the least important predictive factor.