Prediction of Myasthenia Gravis (MG) Crisis Events by a Machine Learning (ML) Algorithm
Author(s)
Nicholas Streicher, MD1, Karen S. Yee, PhD2, Justin Lee, PhD, BS2, Nicholas J Silvestri, MD3;
1MedStar Georgetown University Hospital, UT Neurology, Washington, DC, USA, 2Alexion, AstraZeneca Rare Disease, Boston, MA, USA, 3Jacobs School of Medicine and Biomedical Sciences, University at Buffalo, Buffalo, NY, USA
1MedStar Georgetown University Hospital, UT Neurology, Washington, DC, USA, 2Alexion, AstraZeneca Rare Disease, Boston, MA, USA, 3Jacobs School of Medicine and Biomedical Sciences, University at Buffalo, Buffalo, NY, USA
Presentation Documents
OBJECTIVES: To identify key patient characteristics, disease symptoms, and comorbidities associated with an increased risk of MG crisis using a ML algorithm.
METHODS: Data for patients with and without MG crisis events at year 1 were extracted from the IQVIA Pharmetrics® Plus database (01Jan2015-31Dec2022) and compared; the Optum Clinformatics® database was used for external model validation. Model-identified variables for examination included demographic characteristics and baseline symptoms, comorbidities, treatment, and hospitalization events; highly correlated variables were controlled for multicollinearity. A regularized logistic regression model was applied to the identified variables using the H2O AutoML package.
RESULTS: Patients with MG crises at year 1 (n=332) were significantly (P<0.05) more likely than those without MG crises at year 1 (n=7277) to be older at MG index date (mean±SD: 61.5±14.5 vs 57.1±14.8 years, respectively), have higher rates of comorbidities/symptoms (eg, shortness of breath [21.1% vs 9.3%], dysphagia [20.0% vs 9.0%]), and have certain prescription claims (eg, albuterol sulfate [11.7% vs 4.9%], amlodipine besylate [11.4% vs 5.6%]). Major model-identified MG crisis risk factors included inpatient place of service (odds ratio [OR; 95%CI]: 1.96 [1.38-2.78], P<0.001), aphasia/speech disturbance/dysphagia (1.50 [1.13-1.99], P<0.01), age (1.01 [1.00-1.02], P=0.03), shortness of breath/other dyspnea (1.37 [1.01-1.85], P=0.04), and cough (1.45 [1.04-2.02], P=0.03) at baseline. Survival analyses confirmed significant associations between these risk factors and likelihood of MG crises, with higher probabilities observed for patients with multiple risk factors. Internal validation of the model by receiver operating characteristic curve yielded an area under the curve (AUC) of 0.71 (95%CI: 0.69-0.72), indicating fair overall predictive performance; external model validation yielded a similarly acceptable AUC (0.65).
CONCLUSIONS: This novel analysis using a ML algorithm identified several key patient-relevant characteristics, disease symptoms, and comorbidities that are risk factors for experiencing an MG crisis, and these results may help inform future treatment strategies to reduce the risk of MG crises.
METHODS: Data for patients with and without MG crisis events at year 1 were extracted from the IQVIA Pharmetrics® Plus database (01Jan2015-31Dec2022) and compared; the Optum Clinformatics® database was used for external model validation. Model-identified variables for examination included demographic characteristics and baseline symptoms, comorbidities, treatment, and hospitalization events; highly correlated variables were controlled for multicollinearity. A regularized logistic regression model was applied to the identified variables using the H2O AutoML package.
RESULTS: Patients with MG crises at year 1 (n=332) were significantly (P<0.05) more likely than those without MG crises at year 1 (n=7277) to be older at MG index date (mean±SD: 61.5±14.5 vs 57.1±14.8 years, respectively), have higher rates of comorbidities/symptoms (eg, shortness of breath [21.1% vs 9.3%], dysphagia [20.0% vs 9.0%]), and have certain prescription claims (eg, albuterol sulfate [11.7% vs 4.9%], amlodipine besylate [11.4% vs 5.6%]). Major model-identified MG crisis risk factors included inpatient place of service (odds ratio [OR; 95%CI]: 1.96 [1.38-2.78], P<0.001), aphasia/speech disturbance/dysphagia (1.50 [1.13-1.99], P<0.01), age (1.01 [1.00-1.02], P=0.03), shortness of breath/other dyspnea (1.37 [1.01-1.85], P=0.04), and cough (1.45 [1.04-2.02], P=0.03) at baseline. Survival analyses confirmed significant associations between these risk factors and likelihood of MG crises, with higher probabilities observed for patients with multiple risk factors. Internal validation of the model by receiver operating characteristic curve yielded an area under the curve (AUC) of 0.71 (95%CI: 0.69-0.72), indicating fair overall predictive performance; external model validation yielded a similarly acceptable AUC (0.65).
CONCLUSIONS: This novel analysis using a ML algorithm identified several key patient-relevant characteristics, disease symptoms, and comorbidities that are risk factors for experiencing an MG crisis, and these results may help inform future treatment strategies to reduce the risk of MG crises.
Conference/Value in Health Info
2025-05, ISPOR 2025, Montréal, Quebec, CA
Value in Health, Volume 28, Issue S1
Code
MSR115
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Neurological Disorders