INTRODUCTION
Artificial intelligence (AI) and machine learning techniques have shown promise in individualizing medical treatments. One of the main limitations of machine learning applications in translational research is their complexity and the fact that no clear conclusions can be made for which features are the most defining ones for each prediction.
This analysis reviews an interpretable machine learning method and a user-friendly tool to individualize treatment of chronic pain patients by predicting the benefits of a single treatment: the Salonpas® Pain Relieving Patch.
- Chronic pain adversely affects function and quality of life
- Topical analgesics are effective therapies and offer pain relief devoid of the risks of abuse, misuse, and addiction
- Identifying optimal treatments for pain is difficult due to high variability in patient demographics, clinical profile, adverse effects and analgesic response.
- The ability to predict treatment response would have a tremendous benefit to clinicians and patients.
Artificial intelligence (AI) and machine learning techniques have shown promise in individualizing medical treatments. One of the main limitations of machine learning applications in translational research is their complexity and the fact that no clear conclusions can be made for which features are the most defining ones for each prediction.
This analysis reviews an interpretable machine learning method and a user-friendly tool to individualize treatment of chronic pain patients by predicting the benefits of a single treatment: the Salonpas® Pain Relieving Patch.
METHODS & MATERIALS
Data were evaluated from 186 pain patients enrolled in an Institutional Review Board (IRB) approved study (RELIEF) after use of a topical pain-relieving analgesic patch for 14 days. All patients replied to 41 questions before and 14-days after the patch treatment was employed. These questions were formulated as features by encoding categorical variables to features using the FeatureHasher method to allow for their integration with numeric features, ending up with a list of 85 features.
Data were split into training (n=152) and testing (n=34) patient sets in a stratified manner. Three basic endpoints were examined for the prediction models: total BPI Severity scores, total BPI Interference scores, and changes in the total drugs.
A novel interpretable machine learning method was developed based on a multi-objective ensemble classification/regression technique. (See QR code publication for additional details.
To further evaluate the performance of the trained outcome prediction models for the Salonpas® Pain Relieving Patch into individualizing the therapy of chronic pain patients, we applied the trained models to predict responders and non-responders of the Salonpas® Pain Relieving Patch in the participants of another IRB-approved study consisting of 631 patients.
Data were evaluated from 186 pain patients enrolled in an Institutional Review Board (IRB) approved study (RELIEF) after use of a topical pain-relieving analgesic patch for 14 days. All patients replied to 41 questions before and 14-days after the patch treatment was employed. These questions were formulated as features by encoding categorical variables to features using the FeatureHasher method to allow for their integration with numeric features, ending up with a list of 85 features.
Data were split into training (n=152) and testing (n=34) patient sets in a stratified manner. Three basic endpoints were examined for the prediction models: total BPI Severity scores, total BPI Interference scores, and changes in the total drugs.
A novel interpretable machine learning method was developed based on a multi-objective ensemble classification/regression technique. (See QR code publication for additional details.
To further evaluate the performance of the trained outcome prediction models for the Salonpas® Pain Relieving Patch into individualizing the therapy of chronic pain patients, we applied the trained models to predict responders and non-responders of the Salonpas® Pain Relieving Patch in the participants of another IRB-approved study consisting of 631 patients.
RESULTS
FIGURE 1
Comparative results of trained machine learning models in predicting A. Total Severity Change, B. Total Interference Change and C. Total Drugs Change.
Comparative results of trained machine learning models in predicting A. Total Severity Change, B. Total Interference Change and C. Total Drugs Change.
FIGURE 2
PCA representation of the participants of RELIEF Patients using the selected features from the proposed machine learning method. Clustering was conducted using k-prototypes method
PCA representation of the participants of RELIEF Patients using the selected features from the proposed machine learning method. Clustering was conducted using k-prototypes method
FIGURE 3
Screenshot of web tool for personalizing the application of Salonpas® Patch using the trained machine learning models. After the data of the patient are being stored, the predicted decrease of BPI Severity, BPI Interference and Total Number of Drugs after using Salonpas® Patch are displayed to allow a more informed decision about its suitability with the patient’s profile
Screenshot of web tool for personalizing the application of Salonpas® Patch using the trained machine learning models. After the data of the patient are being stored, the predicted decrease of BPI Severity, BPI Interference and Total Number of Drugs after using Salonpas® Patch are displayed to allow a more informed decision about its suitability with the patient’s profile
Results from the original RELIEF study showed that 89% of participants in the treatment group improved their total severity and total interference scores and during the study 42% reduced the total number of analgesic drugs used.
Our experimental results (Figure 1) demonstrated that the machine learning models were able to predict endpoints with high accuracy, with the AUC exceeding 90% and Spearman correlation metric exceeding 0.4 for all endpoints, far exceeding the test set performance of other benchmark models.
The machine learning method reduced the number of required data point analyses from 85 to 19 and was able to define groups of responders and non-responders of the treatment group.
The Salonpas® Pain Relieving Patch provided maximal benefits in patients who take high number of anti-inflammatory drugs, use Voltaren®, have pain in knee or foot, do little or no heavy physical exercise and have problems with sleep (Figure 2: red cluster).
The treatment group appeared less effective for patients who do not have a confirmed diagnosis of chronic pain and those who already use alternative over the counter medications (Figure 2: dark green cluster).
Single feature correlation analysis demonstrated that while using the Salonpas® Pain Relieving Patch:
A new web-tool (https://relief.insybio.com/) was designed to provide access to the trained prediction models to physicians and chronic pain patients (Figure 3).
Our experimental results (Figure 1) demonstrated that the machine learning models were able to predict endpoints with high accuracy, with the AUC exceeding 90% and Spearman correlation metric exceeding 0.4 for all endpoints, far exceeding the test set performance of other benchmark models.
The machine learning method reduced the number of required data point analyses from 85 to 19 and was able to define groups of responders and non-responders of the treatment group.
The Salonpas® Pain Relieving Patch provided maximal benefits in patients who take high number of anti-inflammatory drugs, use Voltaren®, have pain in knee or foot, do little or no heavy physical exercise and have problems with sleep (Figure 2: red cluster).
The treatment group appeared less effective for patients who do not have a confirmed diagnosis of chronic pain and those who already use alternative over the counter medications (Figure 2: dark green cluster).
Single feature correlation analysis demonstrated that while using the Salonpas® Pain Relieving Patch:
- the higher the initial total interference score, the larger the margin for improvement using RELIEF study data;
- the greater the number of medications, the greater the reduction of medications.
A new web-tool (https://relief.insybio.com/) was designed to provide access to the trained prediction models to physicians and chronic pain patients (Figure 3).
CONCLUSIONS
Topical analgesics can be of benefit too many, but not all pain patients. AI, machine learning, and predictive analytics models can be successfully incorporated into medical decision making. This novel precision medicine approach allows clinicians to select the treatments that have a high probability of success. Historically, these tools had limited generalizability properties restricting their application in clinical practice.
Here we introduced a machine learning technique to predict patients’ response to OTC topical analgesics, overcoming previous limitations of machine learning techniques. This method was able to identify the clinical characteristics associated with a group of “super responders”. The ultimate goal is to prospectively predict which patients are likely to get the most benefit from a study treatment- in this case the Salonpas® Pain Relieving Patch.
This predictive analytic methodology can be applied to separate and larger datasets and used either retrospectively or prospectively, to validate or analyze whether a certain treatment is or would be effective in a given population.
Topical analgesics can be of benefit too many, but not all pain patients. AI, machine learning, and predictive analytics models can be successfully incorporated into medical decision making. This novel precision medicine approach allows clinicians to select the treatments that have a high probability of success. Historically, these tools had limited generalizability properties restricting their application in clinical practice.
Here we introduced a machine learning technique to predict patients’ response to OTC topical analgesics, overcoming previous limitations of machine learning techniques. This method was able to identify the clinical characteristics associated with a group of “super responders”. The ultimate goal is to prospectively predict which patients are likely to get the most benefit from a study treatment- in this case the Salonpas® Pain Relieving Patch.
This predictive analytic methodology can be applied to separate and larger datasets and used either retrospectively or prospectively, to validate or analyze whether a certain treatment is or would be effective in a given population.