Supporting Machine Learning Model in the Treatment of Chronic Pain

Abstract

Conventional therapy options for chronic pain are still insufficient and patients most frequently request alternative medical treatments, such as medical cannabis. Although clinical evidence supports the use of cannabis for pain, very little is known about the efficacy, dosage, administration methods, or side effects of widely used and accessible cannabis products. A possible solution could be given by pharmacogenetics, with the identification of several polymorphic genes that may play a role in the pharmacodynamics and pharmacokinetics of cannabis. Based on these findings, data from patients treated with cannabis and genotyped for several candidate polymorphic genes (single-nucleotide polymorphism: SNP) were collected, integrated, and analyzed through a machine learning (ML) model to demonstrate that the reduction in pain intensity is closely related to gene polymorphisms. Starting from the patient's data collected, the method supports the therapeutic process, avoiding ineffective results or the occurrence of side effects. Our findings suggest that ML prediction has the potential to positively influence clinical pharmacogenomics and facilitate the translation of a patient's genomic profile into useful therapeutic knowledge.

Keywords: cannabis; machine learning; pain treatment; pharmacogenetics; precision medicine.

Figure 1

Features related to each patient, including static features in light yellow and dynamic features over the follow-up (F/U) periods in light green. In dark green, a focus on the dynamic features of each subperiod is shown.

Figure 2

Medical cannabis dose distribution range. Boxplot of CBD and THC mg/day distribution. Lower and upper box boundaries represent the 25th and 75th percentiles, respectively; the red line inside the box shows the median; and the lower and upper error lines highlight the 10th and 90th percentiles, respectively. Outliers’ data falling outside the 10th and 90th percentiles were not reported.

Figure 3

Patients who decided to attend all the follow-up visits were completely described by four rows. Drop-out case examples include a patient who dropped the treatment after the second follow-up visit, represented therefore by only two rows.

Figure 4

Real observations are shown as single red points within the plot, while predicted values are represented as single blue points. THC and CBD mg/day predicted vs. real values are plotted on the left and the right side of the plot, respectively. In the X-axis, 20 values for each cannabinoid are reported. The Y-axis represents daily dose values. The lower graph includes error bars for each prediction, which represents the variation of the corresponding coordinate of the point.

Figure 5

Feature importance bar graph for THC and CBD daily dose prediction on the left and on the right, respectively. All the features are listed and sorted by their importance. The X-axis displays feature names used as the input of the ML model, while the Y-axis show which features attribute the most predictive power to the model.