Pharma R&D Today
Ideas and Insight supporting all stages of Drug Discovery & Development
Using Machine Learning for Identifying and Classifying Patient Groups
Posted on August 14th, 2017 by Dr Andrew A. Parsons in Pharma R&D
To maximize value to payers, patients, and practitioners, pharma R&D wants to ensure that new medicines are given to the right patient, at the right time and at the right dose. To do this appropriately, we need ways to classify or diagnose patient groups.
Over the last decade, significant progress has been made, notably in oncology—but significant issues remain with some neurological and psychiatric conditions.
With this challenge in mind, news articles caught my attention this week. It focused on the idea that machine learning could lead to objective measures for a diagnosis of schizophrenia (1). In this case, an objective measure would allow a prospective and independent diagnosis based on brain function, rather than by behavioral assessment and exclusion of other conditions. Machine learning could play a significant role in creating independent measures and analysis following some simple sets of rules. Gheiratmand and colleagues (2) used an open source dataset to develop their comparison involving correlations of multi-functional networks at various degrees of resolution. Using BOLD, a measure of blood oxygenation as a marker of neuronal activity, they assessed brain activity during recognition of dissimilar auditory stimuli.
Their findings support hyper-connectivity across brain areas (particularly across fronton-parietal regions) and demonstrate differences between individuals with symptoms of schizophrenia and healthy controls. They also demonstrate that activity in certain functional networks predicted clinical scores on several measures. What seemed interesting to me was that the functional brain networks that were most predictive of clinical scores were not the ones that were most highly altered in schizophrenia. The level of accuracy was good at 74% for using the multivariate comparisons to predict clinical severity. However, as a patient or clinical practitioner, improved levels of accuracy would be required for a definitive diagnosis.
Other studies indicate high accuracy levels are achievable. Machine learning approaches using data from structural MRI of gray matter density have been used to distinguish between healthy controls, schizophrenics, and patients with bipolar disorder. Schizophrenia patients could be determined with an accuracy of 90% from healthy controls and 88% of patients with bipolar disorder (3). This level of performance can certainly add value and understanding to all involved.
There seems to be a growing expertise in the use of analytical technologies to support evidence-based decision making in psychiatry, with companies developing products to create objective measures of brain function (1). The scope for linking behavior to function could have tremendous value in the earlier diagnosis of dementia, creating the ability for more accurate diagnosis of neurological and psychiatric disorders. Measuring, classifying and diagnosing are hallmarks of the medicines development process and ways to identify common pathways and patient groups will aid pharmaceutical development in mental health. These approaches also need to be balanced with the views of the patient.
Patient-focused endpoints are gaining increasing importance and so should individual differences and the diversity we have amongst the human condition. There is a risk that such precise classification will disempower the patient. For example, I recall a conversation at a scientific meeting some years ago, where the researcher proclaimed that they were developing a “biomarker for happiness”. A comment and question came from the floor: “I can’t see the value in this, can’t you just ask them?” To which the researcher answered, “Happiness is such a wooly concept that can change rapidly. With a biomarker, I will know they are happy”. An expert mindset from a well-meaning professional has the potential to disempower and marginalize the experience of the individual, locking them into a classification that may not be real for them. As our behaviors and actions become increasingly analyzed by wearable technology, we will become increasingly categorized into groups.
No doubt new technologies will have a tremendous impact in both the diagnosis and early treatment of neuropathology. This area is developing rapidly in many directions. From a more humanistic perspective, we also need to understand how technology, in its attempts to create objectivity, will impact the subjective sense of who we are and our ability to manage our situation (our self-efficacy). These factors are key contributors to health and well-being and remind us that we need to keep the focus on the patient.
Objective classifications will be most useful when we understand how they will be used subjectively by patients, payers, and practitioners.
- Gheiratmand et al. Learning stable and predictive network-based patterns of schizophrenia and its clinical symptoms. 2017. Schizophrenia 3: 22. Doi:10.1038/s41537-017-0022-8
- Schnack et al. Can structural MRI aid clinical classification? A machine learning study in two independent samples of patients with schizophrenia, bipolar disorder, and healthy subjects. 2014. Neuroimage 84: 299-306
Access the complimentary white paper ‘Precision Medicine: The new R&D paradigm‘
All opinions shared in this post are the author’s own.
R&D Solutions for Pharma & Life SciencesWe're happy to discuss your needs and show you how Elsevier's Solution can help.
Dr Andrew A. Parsons
Director of Reciprocal Minds Limited & Chairman of Pharmasum Therapeutics AS
- Hopes, Issues, and Concerns in Tackling the Challenge of Dementia
- Start-Up Spotlight: Innovative NeuroTechnologies, Inc.
- Let’s focus on the system