IDENTIFYING ADOLESCENTS AT RISK OF DEVELOPING NEGATIVE OUTCOMES AFTER RECEIVING OPIOID ANALGESICS FOR CHRONIC NON-CANCER PAIN MANAGEMENT USING MACHINE LEARNING ALGORITHMS

Canada's prescription opioid dispensing rates have increased since the early 21st century and this has contributed to an increase in opioid-related morbidity and mortality. Adolescents are one of the most vulnerable age groups when it comes to experiencing morbidity and mortality related to opioids. Adolescents who consume prescription opioids are also more susceptible to developing substance use disorder in later years. This thesis has two goals. The first goal is to study the epidemiology of prescription opioid dispensation among adolescents aged 12 to 17 years residing in Alberta between April 1st, 2010, and March 31st, 2015 who were treated for non-cancer pain. Its second goal is to create a supervised machine learning model that could predict the occurrence of negative outcomes following the dispensation of prescription opioids for chronic non-cancer pain management among the above-mentioned study population. This study relied on the administrative data gathered by Alberta Health, particularly the use of the community pharmacies dataset. In order to achieve the goals of this research it was necessary to construct episodes of opioid consumption for study subjects. All incidences of opioid dispensations to study subjects were found and supply day quantities plus a 60-day washout were added to them to determine the predicted end date of each dispensation. The occurrence of a successive opioid dispensation that happened before the predicted end date of the last dispensation would constitute a new event in that episode. In total 78805 opioid prescriptions were filled by study subjects during this study period of which %82.17 were incidence cases. The incidence proportion of opioid-containing dispensations among the study population had an increasing trend over the study period. Furthermore, male subjects and rural residents had a higher ratio of opioids dispensed overall dispensations for adolescent males and rural residents compared to the same ratios for females and urban residents. Based on the left-skewed age histogram of opioid dispensations older adolescents were dispensed more opioids than younger ones. An advantage of supervised machine learning algorithms is their ability to make predictions about unseen data by generalizing from observed pieces of evidence. To achieve the second goal of this research, prescription opioid dispensation episodes lasting 90 days or longer were extracted in order to remove episodes that were acute pain related. Overall 699 eligible episodes were deemed eligible as chronic non-cancer pain therapies of which 71 episodes had resulted in negative outcomes up to one year after the end of the episode. Among all the features that were considered in this study, duration and number of dispensations in an episode, subject’s age at baseline, and having a mental disease diagnosis at the baseline were significantly different among those with and without negative outcomes. A random forest model with accuracy= 0.71, AUC= 0.85, recall= 0.86, precision= 0.61, and f-score= 0.72 was superior to other models. This algorithm was particularly interesting as it used only three features for its predictions; episode length, number of dispensations in the episode, and mental disorder diagnoses at baseline. While opioid dispensation to most age groups in Canada has had decreasing trends in recent years, the incidence proportion of prescription opioid dispensation for non-cancer pain to adolescents 12 to 17 years of age in Alberta had an increasing trend from 2010 to 2015. However, it is possible to create a simple machine learning algorithm with very few features that can predict, with good sensitivity, which episodes of opioid dispensation for chronic con-cancer pain among individuals aged 12 to 17 will result in experiencing negative outcomes. These are concerning and promising results for prescribers and policy-makers.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.