Preliminary Results of a Machine Learning Model for the Detection of Undiagnosed Cases of Ankylosing Spondylitis in Italy

OBJECTIVES: Rapidly evolving and disabling diseases might significantly benefit from early diagnosis to decrease the number of undiagnosed cases and to allow prompt interventions. This is a preliminary model of a synergistic utilization of machine learning (ML) and patient healthcare data available in administrative databases in the setting of Ankylosing Spondylitis (AS) to be used as a forerunner template for other diseases where a timely detection of undiagnosed cases might be critical in clinical decision-making.

METHODS: A dedicated algorithm was developed using data from administrative databases of Italian healthcare entities covering about 12 million health-assisted subjects. Drug prescriptions, hospitalization discharge diagnoses and exemption codes extracted from Jan-2009 to Sep-2023 were the input variables to deploy two ML models, based on a neural network approach (a system to mimic the human brain) and a random forest algorithm to provide a prediction model as output.

RESULTS: On a population of AS patients and randomly selected non-AS controls of equal size (N=7,934), ML allowed to properly discriminate affected and non-affected subjects, with good accuracy scores (% correct predictions of total predictions), reaching 79.3% and 80.2% for neural network and random forest models, respectively. Using this approach, the strongest predictors of AS diagnosis were resulted to be the following: prescriptions for autoimmune disease treatment, hospitalization code ICD-9-CM 724 (other and unspecified back disorders), and exemption code 009 (detecting inflammatory bowel diseases).

CONCLUSIONS: The present model describes a successful joint utilization in the setting of AS of ML and real-world data from a sample corresponding to nearly 20% of the Italian population. This pivotal design might pave the way for further clinical applications, providing a valuable tool to facilitate the early detection of undiagnosed cases of special importance for rapidly evolving and disabling diseases.