How Aivo Health uses Probabilistic Machine Learning to revolutionize chronic pain treatment

Why is Machine Learning needed to treat chronic pain?

Efficient and effective chronic pain treatment is a huge challenge throughout the entire healthcare industry – you could even say that chronic pain care is broken. At Aivo Health, we are fixing chronic pain treatment by knocking down the barriers that typically impede standard forms of care:

• Complexity: Chronic pain is a multifaceted phenomenon driven by biological, psychological, and social factors that vary between people. Rapid and reliable measurement of the factors underlying this complexity is required to fully understand pain.

• Personalization: Consistent data-driven, patient-centric decisions must be made to provide the most optimal treatment. In standard primary care settings, most patients feel that they are undertreated. New scientific evidence suggests that personality traits are linked to how patients experience chronic pain. Understanding these traits is central for building an effective treatment.

• Accessibility: Multimodal treatments are the gold standard for chronic pain. However, such intensive treatments cannot be made available for the majority of patients. Even when available, waitlists are often months long, which can result in significant deterioration of health and quality of life. There are simply not enough clinicians available to care for the growing number of people with chronic pain.

How can Machine Learning help?

Probabilistic Machine Learning (probabilistic ML; also referred to as Bayesian Machine Learning) is the practice of building models that represent uncertainty using probability distributions. With modern probabilistic programming languages such as Stan, we are able to fit models with thousands of parameters with relative ease.

At Aivo Health we develop probabilistic ML models that represent expert clinical domain knowledge in a way that traditional ML and artificial intelligence (AI) models can not. This results in human-interpretable models that provide useful and robust insights even when little data is available per patient.

Since the models we build are all custom-made, they are 100% explainable. This is critical in the healthcare context. Nobody wants health-related decisions from a black box—being able to explain why a certain treatment is likely to work is important for both transparency and compliance.

Altogether, our probabilistic ML models help us surmount the challenges of chronic pain treatment in the following ways:

• Complexity is embraced: We used our models to comb through tens of thousands of questionnaire responses from several hundred chronic pain patients to identify the most relevant factors underlying people’s experience with chronic pain. This work enables us to estimate a set of “Recovery Factors” for each person by having them respond to only 30 questions. A process that once took up to an hour now takes a few minutes, with minimal loss of information regarding the complexity of a person’s experience with pain. This allows us to provide personalized treatment from day 1.

• Personalization is calibrated: A major benefit of probabilistic ML is that all information we extract from a model comes along with uncertainty quantification. Whereas most traditional ML/AI models make point predictions, probabilistic models provide true uncertainty around the predictions. This allows us to make more ethical, patient-centric treatment decisions (see Image below).

‍

• Accessibility is enhanced: Probabilistic ML allows for us to capture all the information we need to kickstart someone’s pain relief journey with a brief 30-item questionnaire on day 1, all from the comfort of their own home. Our digital clinic at Aivo Health therefore allows for anyone to access our services at the very moment they are in need—no waitlists, no transportation barriers, and no scheduling conflicts. 

How is Machine Learning used in context?

We use probabilistic ML throughout our members’ journey in the Aivo Health program—from the in-app assessment on the first day to the specific treatment recommendations we make in weeks and months to come. 

Our completely automated workflow begins by passing in-app question responses through a dimensionality reduction model, which allows us to extract information on the various treatment factors that underlie each member’s unique form of pain (we term these “Recovery Factors”). We then use these Recovery Factors to recommend specific treatment “tracks”, each which takes 3 to 5 days to complete. Throughout, members enter their daily pain, mood, and sleep ratings, which we then use to continuously fine-tune track recommendations (see image below for a workflow diagram, and our accompanying member-focused blog for more information). 

Although it is automated, we ensure that all aspects of our ML workflow are transparent and explainable. For example, human Health Coaches have the final say over all track recommendations sent to members. Additionally, our data science team is up to date with state-of-the-art ML-Ops practices that allow for us to continuously monitor how and why our ML models make specific recommendations.

‍