Machine Learning-Based Approach to Discover Subtypes of Light-Chain Amyloidosis Using US Claims Data

OBJECTIVES: Light-chain Amyloidosis (AL) is a rare disease caused by plasma cell dyscrasia that affects multiple organ systems. This study aimed to use machine learning techniques to discover AL subtypes based on the organ systems impacted and to assess the prognostic effect of the patient subtypes on the outcomes following diagnosis.

METHODS: We used IQVIA PharMetrics^® Plus data from Jan-2016 to Sept-2022 with the following inclusion criteria: ≥2 AL diagnosis codes (ICD-10:E85.81), ≥24 months of continuous enrollment pre-diagnosis, and ≥6 months of continuous enrollment post-diagnosis. We identified AL patient’s clinical manifestations pre-diagnosis, including cardiovascular, renal, hepatic, neurologic, and others. We trained an unsupervised machine learning algorithm, K-Modes, to cluster the patients into subtypes based on their manifestations. K-Modes accounts for correlation between manifestations and discovers clusters based on similarity between patients. Discovered clusters/subtypes were labelled according to the prevalent contributing manifestations within each. For each subtype, we measured the probability of an emergency room (ER) visit or an inpatient (IP) stay in the 6 months post-diagnosis and calculated the odds ratio (OR) using logistic regression, adjusting for age, sex, and region.

RESULTS: Among N=1276 AL patients meeting our inclusion criteria, we discovered the following subtypes: Cardiac (N=200,16%), characterized by cardiomyopathy and heart failure without edema; Severe Cardiac (N=116,9%) characterized by heart failure with edema; Renal (N=253,20%), characterized by proteinuria, chronic kidney disease; Cardiac+Renal (N=96,8%), characterized by heart failure with CKD, frequently comorbid with neuropathy; Non-Specific (N=135,11%), characterized by GI/hepatic manifestations; and Early Mayo Stage (N=476,37%), characterized by a lack of manifestations. Regression analysis revealed ~12x higher odds of a Cardiac ER Visit in the Severe Cardiac subtype compared to Early Mayo Stage subtype (OR=12.2,95%CI=6.5,23.7;p<0.001), with similar results for IP stays.

CONCLUSIONS: Utilizing machine learning can be useful in effectively identifying patient subtypes for AL, which can impact treatment regimens to control symptoms.