Transforming Healthcare Analytics with Machine Learning

166

By Zeynep Icten, Ph.D., Director of Data Science Solutions at Panalgo

The digitization of the healthcare industry has provided analytics and life sciences professionals with more data than ever before. Wearable fitness trackers, electronic health records, and remote patient monitoring tools now provide a broad array of insights about patient populations and behaviors that can be used for initiatives including product development and launch, determining unmet needs, and predicting patient outcomes, among others. 

While this deluge of data is advantageous for improving the healthcare system and patient outcomes, it requires more sophisticated interpretation and analysis. Here, there is an opportunity to not only test hypotheses, but to use algorithms to search for patterns – which can lead to novel insights that might otherwise have been overlooked.

How is Machine Learning Different from Other Forms of Analysis?

As the abundance and accessibility of available data continues to revolutionize the healthcare industry, data analysts and other key stakeholders need the appropriate tools to synthesize this information and harvest useful insights. Machine learning is an area of artificial intelligence (AI) consisting of a collection of methodologies that focus on algorithmically learning efficient representations of data and extracting insights. Unlike inference-focused approaches, machine learning methods can be used to learn from complex data sets and to discover patterns that traditional statistical inquiries may not uncover. 

In healthcare, machine learning can be used to efficiently analyze broad collections of data to uncover patterns in, for example, which patient populations might benefit from certain treatments or interventions. Machine learning models can also predict a wide range of outcomes such as a patient’s hospital readmission risk and potential for disease recurrence. 

Machine learning is not a magic bullet – it doesn’t replace studies that seek to test specific hypotheses. Rather, it complements these methods, expanding upon previous findings from traditional analyses. It adeptly manages covariates and high-dimensional data, providing the specific and sensitive algorithms needed to uncover relationships between data elements that can be highly nonlinear and complex.

Using Machine Learning to Predict Patient Outcomes

By leveraging machine learning as a predictive tool, providers and payers can better identify high-risk patients to intervene prior to rehospitalization. Consider this: There are nearly a million people in the U.S. living with multiple sclerosis (MS), many of whom experience relapses, which are associated with disability progression and worsening outcomes. With a greater understanding of the underlying causes of relapse, healthcare professionals can mitigate more significant impacts and improve the management of this disease. 

To do just this, my colleagues and I recently studied administrative claims data among MS patients to identify predictors of inpatient relapse. By analyzing real world data (RWD) with machine learning approaches, we were able to develop robust, strong predictive models, creating a data-driven decision rule to discriminate between MS patients with and without an inpatient relapse.

With these models, we achieved an area under the curve (AUC) of 79.3%, sensitivity of 69%, and specificity of 75% using predictors of MS relapse. These predictors include previous inpatient or emergency room visits with an MS diagnosis, the number of MS-related encounters, the number of comorbidities, the use of home care services and durable medical equipment, epilepsy or convulsions, paralysis, urinary tract infections, and the use of muscle relaxants, anticonvulsants and antidepressants. We were also able to identify notable factors protective against relapses. 

Ultimately, we defined a decision rule indicating that patients were more likely to have a relapse if they have 30 or more unique comorbidities or have a previous emergency room visit with an MS diagnosis and 10 or more previous MS related encounters or have 20 or more previous MS related encounters. After appropriate external validation, our findings could potentially be leveraged at point of care to intervene ahead of potential relapses by identifying at-risk patients that may benefit from additional care, like efforts to increase medication adherence or a new medication regimen. 

This decision rule could be used as a proxy for disease severity in database studies using other datasets to stratify patients based on their likelihood of relapse. Ultimately, these actions can potentially improve patients’ quality of life and reduce the need for additional touchpoints with the healthcare system, saving both patients and physicians time and money.