How deep learning can help predict disability progression in multiple sclerosis
Multiple sclerosis (MS) is a chronic neurological disease that affects millions of people worldwide. MS causes inflammation and damage to the protective coating of nerve fibers, leading to various symptoms such as vision problems, fatigue, numbness, and difficulty walking. MS can be classified into two main types: relapsing-remitting MS (RRMS), which is characterized by episodes of inflammation followed by periods of recovery, and progressive MS (PMS), which is marked by continuous worsening of disability without clear relapses.
There is currently no cure for MS, but there are several treatments that can reduce the frequency and severity of relapses, slow down the accumulation of disability, and improve the quality of life of patients. However, not all patients respond equally to these treatments, and some may experience adverse effects or lack of efficacy. Moreover, there is a lack of reliable biomarkers that can predict how a patient will respond to a given treatment, especially for disability progression, which is the most important outcome for patients and clinicians.
To address this challenge, a team of researchers from Canada and Italy have developed a novel method that uses deep learning to estimate the individual treatment effect on disability progression in MS patients. Deep learning is a branch of artificial intelligence that uses complex mathematical models called neural networks to learn from large amounts of data and make predictions or decisions. The researchers used a type of neural network called a multi-headed multilayer perceptron (MLP) to analyze data from six randomized clinical trials involving 3,830 MS patients who received different treatments or placebo.
The MLP was trained to estimate the conditional average treatment effect (CATE), which is the difference in the probability of disability progression between two treatment options for a given patient. The CATE was calculated using baseline clinical and imaging features, such as age, sex, disease duration, disability level, brain volume, and lesion load. The MLP was first pre-trained on a subset of RRMS patients (2,520), then fine-tuned on a subset of primary progressive MS (PPMS) patients (695).
The researchers then tested the performance of the MLP on two separate held-out test sets of PPMS patients who received either anti-CD20 antibodies or placebo (297) or laquinimod or placebo (318). Anti-CD20 antibodies are monoclonal antibodies that target a protein on the surface of B cells, which are immune cells involved in inflammation and autoimmunity. Laquinimod is a small molecule that modulates the immune system and reduces inflammation and neurodegeneration.
The results showed that the MLP could identify patients who were more likely to benefit from anti-CD20 antibodies or laquinimod compared to placebo. The average treatment effect was larger for the 50% and 30% of patients predicted to be most responsive by the MLP than for the entire group. For example, for anti-CD20 antibodies, the hazard ratio (HR) for disability progression was 0.492 (95% confidence interval [CI], 0.266-0.912; p = 0.0218) for the top 50% and 0.361 (95% CI, 0.165-0.79; p = 0.008) for the top 30%, compared to 0.743 (95% CI, 0.482-1.15; p = 0.179) for the whole group.
The researchers also showed that using the MLP for predictive enrichment could increase the statistical power of clinical trials by preferentially randomizing patients who were more likely to respond to treatment. This could enable shorter and smaller proof-of-concept trials with disability progression as an endpoint, which could accelerate drug development and reduce costs.
The study, published in Nature Communications on September 26th, 2022 1, is the first to use deep learning to estimate individual treatment effect on disability progression in MS using baseline features. The authors acknowledge some limitations of their method, such as the need for validation in prospective studies, the dependence on data availability and quality, and the potential ethical and regulatory implications of predictive enrichment. However, they also highlight the potential benefits of their approach for personalized medicine and precision neurology in MS and other neurological diseases.