Multiple Sclerosis(MS)is a disease that disrupts the flow of information within the brain.It affects approximately 1 million people in the US.And remains incurable.MS treatments can cause side effects and impact the q...Multiple Sclerosis(MS)is a disease that disrupts the flow of information within the brain.It affects approximately 1 million people in the US.And remains incurable.MS treatments can cause side effects and impact the quality of life and even survival rates.Based on existing research studies,we investigate the risks and benefits of three treatment options based on methylprednisolone(a corticosteroid hormone medication)prescribed in(1)high-dose,(2)low-dose,or(3)no treatment.The study currently prescribes one treatment to all patients as it has been proven to be the most effective on average.We aim to develop a personalized approach by building machine learningmodels and testing their sensitivity against changes in the data.We first developed an unsupervised predictive-prescriptive model based on K-means clustering in addition to three predictive models.We then assessed the models’performance with patient data perturbations and finally developed a robust model by re-training on a set that includes perturbations.These increased themodels’robustness in highly perturbed scenarios(+10%accuracy)while having no cost in scenarios without perturbations.We conclude by discussing the trade-off between robustification and its interpretability cost.展开更多
文摘Multiple Sclerosis(MS)is a disease that disrupts the flow of information within the brain.It affects approximately 1 million people in the US.And remains incurable.MS treatments can cause side effects and impact the quality of life and even survival rates.Based on existing research studies,we investigate the risks and benefits of three treatment options based on methylprednisolone(a corticosteroid hormone medication)prescribed in(1)high-dose,(2)low-dose,or(3)no treatment.The study currently prescribes one treatment to all patients as it has been proven to be the most effective on average.We aim to develop a personalized approach by building machine learningmodels and testing their sensitivity against changes in the data.We first developed an unsupervised predictive-prescriptive model based on K-means clustering in addition to three predictive models.We then assessed the models’performance with patient data perturbations and finally developed a robust model by re-training on a set that includes perturbations.These increased themodels’robustness in highly perturbed scenarios(+10%accuracy)while having no cost in scenarios without perturbations.We conclude by discussing the trade-off between robustification and its interpretability cost.
