JMIR Publications

ABSTRACT

Background:

With the availability of newer therapies, the duration of therapy (DoT) shortens with each increasing line of treatment in patients with multiple myeloma (MM) in Japan.

Objective:

To identify factors that shorten DoT in MM patients using machine learning (ML) procedure from the Medical Data Vision (MDV) database.

Methods:

This nationwide, retrospective observational cohort real-world study was conducted using anonymized patient data from MDV claims database from 2003-2022. Patients (≥18 years) with transplant-ineligible newly-diagnosed MM (continued 1st line [1L] therapy), or relapsed/refractory MM (continued 2L/3L therapy) were included. To identify important predictive factors, an explainable deep-learning model was created using 647 extracted variables (continuous, binary, and nominal categorical) from MDV database, and the extracted data were used to train ML algorithms to build point-wise linear (PWL) models for predicting DoT. The predictive performance of the PWL model was compared with elastic net (regularized logistic regression) and the XGBoost (boosting trees) models and calculated by area under the curve (AUC) and evaluated by 10-fold double cross-validation. A clustering analysis (k-means method) of 4,848 individual samples was performed to understand the relationship between each sample and DoT (3, 6, and 12 months). The characteristics of clusters and the features of samples belonging to each cluster during and after treatment were studied using correlation analysis.

Results:

Overall, 2,762 (4,848 individual samples) patients were evaluated; mean age: 69.6 years with 52.5% male. The AUC score of the PWL model to predict DoT at 3, 6, and 12 months was 0.61, 0.64, and 0.66, respectively. Based on similarity of coefficients of regression models, samples were categorized into two clusters (cluster A and cluster B) at DoT of 3 months, three clusters (cluster A, cluster B, and cluster C) at 6 months and 12 months (cluster A, cluster B, and cluster C). Cluster B vs cluster A (at 3 months) and cluster C vs cluster A and B (at 6 and 12 months) had a significantly (P<0.01) higher pre-treatment Charlson Comorbidity Index. Furthermore, they also showed lower median of prediction probability. At 3 months in cluster B and at 6 and 12 months in cluster C, the use of immunomodulatory drugs (IMiDs) in treatment for MM was significantly higher in patients who met predicted DoT at each threshold versus the ones who did not. Additionally, use of aspirin was significantly higher in cluster B and cluster C at 3 and 6 months, respectively.

Conclusions:

Applying ML techniques using PWL model yielded efficient results to understand trends associated with treatment and characteristics of Japanese patients with MM whose DoT were shortened. The study demonstrated that patient’s disease status and management related factors including use of IMiDs and management of thromboprophylaxis may be associated with DoT length.