Our pharmaceutical client had a large cohort of healthy patients and patients with a neurological disease. They had clinical and biosensor data from this cohort including: ECG, EEG, and eye tracking in response to social tasks. These participants were not asked to take a drug as a part of this research. The same data was collected for a smaller cohort of patients with a neurological disease, half of which were offered a placebo and half were offered a drug. The goal was to use the large drug-free cohort to determine biomarkers for this neurological disease which differentiated it from healthy individuals, as well as to identify clinically meaningful subtypes.
Heuristic and statistical modeling identified biomarkers that were associated with the neurological disease versus healthy individuals within the drug-free cohort. Our team was able to identify biomarkers that were associated with specific symptom profiles within the neurological disease using spectral clustering methods. These biomarkers were then used as inputs in machine learning models to train a classifier on the neurological disease versus the healthy individuals using the drug-free cohort and predict which patients had a similar biomarker profile within the clinical drug-trial patients. By removing the drug-trial patients who were classified as healthy from analysis, the treatment group showed greater impact on symptom endpoints, while placebo group showed less.