IBM Research and The Michael J. Fox Foundation develop modeling methodology to help understand Parkinson’s Disease using machine learning
Despite progress on many fronts in the management of chronic diseases, there are still many open questions in the field. A better understanding of chronic disease could enable improved patient care management, and faster, more efficient drug development as a result of better designed clinical trials.1 IBM Research is using machine learning as a tool in the pursuit of revealing the complexities of these diseases. At this week’s Machine Learning for Healthcare Conference, we present our progress in this area as motivated by Parkinson’s disease (PD).
PD is a chronic, progressive neurodegenerative disorder with heterogeneous symptoms which may affect both motor and non-motor function. PD is one of the top ten leading causes of death in persons over 65 years of age in the United States.2 It is estimated that 6 million people worldwide and 1 million people in the United States live with PD, and this prevalence is expected to double by 2040 – making the need for research and a better understanding of the disease even more urgent.3
In collaboration with The Michael J. Fox Foundation for Parkinson’s Research, our team of researchers at IBM is aiming to develop improved disease progression models that can help clinicians understand how the disease progresses in relation to the emergence of symptoms, even when those patients are taking symptom-modifying medications.
Time series and forecasting models (techniques used to predict future outcomes based on previously observed data) are not unique to healthcare applications, but their use in healthcare applications can have particular challenges. In this project, the specific goal was to learn clinical states of PD and their corresponding progression, such that clinicians and patients could have a quantitative measure of an individual’s disease.
Disease states for chronic conditions can be thought of in a manner similar to cancer stages, with the important caveat that in the case of chronic conditions, the connection from disease states to biological mechanisms is often not understood. The underlying biology of PD is not yet fully characterized, which makes learning and classifying stages very difficult. However, disease states are still useful as they provide a concise summary of potential motor and non-motor symptoms and allow clinicians to develop targeted therapy and treatment plans. In the future, progression models such as ours may help support various clinical applications. For example, once the disease states are learned, clinicians could quantitatively group patients as well as better predict progression – which could potentially help to inform more personalized patient care and management, as well as more effective drug trials.
In general, progression is not straightforward nor easy to define in Parkinson’s disease. The symptoms and symptom trajectories of PD manifest in a wide range across patients. This makes it difficult for clinicians to definitively and quantitatively assess where an individual may be at a certain point in time, and how advanced their PD truly is.
This issue is made even more challenging due to comparatively small datasets and confounding of the data due to the effects of medications. Medication use presents an interesting challenge in many chronic diseases where drugs are used to alleviate symptoms, but are not thought to alter the course of the disease. For instance, consider a case of an individual with PD experiencing better motor control due to medication, which could lead to the masking of a more advanced disease state.4
Furthermore, patients’ responses to medication may not be consistent across the population, motivating a need for personalizing predictions. For example, one patient’s tremor symptoms may be very responsive to medication, while another could be experiencing less relief from medication, even though their diseases are equally progressed. Many progression models for PD do not account for medication effects.
Identifying these challenges is key to the success of machine learning in healthcare and is an important aspect of our collaboration with The Michael J. Fox Foundation (MJFF). In addition to providing important domain expertise to inform the model requirements, MJFF has funded the collection of one of the most comprehensive and robust volumes of data from individuals with PD – the Parkinson’s Progression Markers Initiative – making it possible to attempt to elucidate these complexities with the help of machine learning.
Our proposed approach addresses the needs of learning disease states while modeling medication effects, which can be a function of disease state and/or personalized response. It is unique in its focus on modeling data from patients who are on medications. The approach builds on the framework of a hidden Markov model and uses variational inference to learn personalized effects. After learning the model, insights can be derived both from interpreting the parameters of the model to learn more about the disease, as well as analyzing predictions for a particular cohort of patients.
The proposed model and learning algorithm will be presented at the 2020 Conference for Machine Learning for Healthcare. Although the work was motivated by PD, we hope it might be useful or inspire similar work and exploration in other chronic conditions such as diabetes, Alzheimer’s disease, and ALS. Developing the tools for analysis is only the first step in the collaboration with The Michael J. Fox Foundation. Our next results will focus on the clinical insights we have derived by applying these models to the extensive amounts of data collected by The Michael J. Fox Foundation’s landmark clinical study, the Parkinson’s Progression Markers Initiative.
*Note: Statements regarding IBM’s future direction and intent are subject to change or withdrawal without notice, and represent goals and objectives only.