New Director of Analytics
I would like to introduce myself as the new Director of Analytics for Motus Nova. This means I am responsible for finding meaning in data that can help decision-making. In the context of stroke rehabilitation, this amounts to utilizing historical data to provide better care.
I received my PhD in Mathematics from Georgia State University in Atlanta, and before joining the team at Motus Nova, I was a researcher in the Department of Biomedical Informatics (BMI) at Emory University. The Emory BMI uses machine learning, among other informatics approaches, to solve healthcare problems including the detection, diagnosis, and treatment of disease.
Machine Learning amounts to developing mathematical model that learns to perform specific tasks without human direction. Essentially, mathematical models are able to identify patterns in high dimensional datasets that have a lot of data points. The simplest machine learning model is a line of best fit with one independent variable (feature) that learns to predict a single dependent variable. For example, if you have a set of athlete heights and weights, you can fit a line to this data (as height increases, so does weight). Then, if you meet a new athlete, you can predict their weight using the line that you fit to the data.
Athlete weights (in kilograms) as a function of their height (in meters) with the line of best fit. Data is taken from a subset of the athlete data from the 2016 Olympics in Rio, courtesy of https://github.com/flother/rio2016.
This model can quickly become more predictive by including features such as gender, sport, and age. The trade-off for better predictions is that the data is harder to visualize, and the models themselves become more difficult to interpret.
In the critical care setting, you have access to a lot more than one variable in your data set, though. Vital signs are monitored continuously and, depending on how sick a patient is, labs are ordered regularly. In the height/weight example it is pretty easy to see that there is an underlying relationship between the two variables by eye. In fact, humans are pretty good at detecting patterns in two and three dimensions. Beyond three dimensions, especially when the variables also change in time, we are not so adept. However, the mathematical concepts behind fitting a function (like a line) to data extend quite nicely to datasets with many variables.
Critical care provides two distinct challenges when it comes to data analysis and modeling. The first is the sensitive nature of the data being collected and used in model development. This reinforces the mentality of being protective with data, and anonymizing it whenever possible. The second is the serious nature of the conditions that place patients in the ICU. It would be an ethical issue to use any mathematical model in place of clinical intervention, so models must be developed closely with clinicians, and be subjected to strict clinical oversight when deployed in the ICU.
My work at Emory focused on developing a machine learning algorithm that could provide objective, autonomous treatment recommendations for sepsis patients. Sepsis is life-threatening condition that arises when the body’s response to infection injures its own tissues. As the damage from the dysregulated immune response gets worse, septic patients have problems maintaining normal blood pressure and require clinical intervention to correct this. To counteract the drop in blood pressure, the two primary forms of intervention are fluids, which raises blood pressure by increasing circulating volume, and vasopressors, which increase blood pressure by constricting the blood vessels and increasing the contractility of the heart. The algorithm studies historical patient data and the corresponding clinical interventions to develop its own treatment policy that is free of the biases that are inevitable (for better or worse) in clinical practice.
Machine learning model (RL Agent) recommendations compared to the clinical actions for a hypotensive sepsis patient. The top two panels provide a comparison between the normalized vasopressor dose (normalized against norepinephrine) and fluid dose (in mL). The RL Agent recommendation does not have memory of clinician actions, so it’s recommendations should be viewed like a GPS suggesting your next direction from your current location, not your historical route. The third panel shows the patient’s mean arterial pressure, and the fourth depicts an idea of how well the patient is doing at the current hour.
While this type of technology is nowhere near replacing ICU doctors, there are real, meaningful benefits to introducing machine learning algorithms to assist clinicians at the bedside.
Stroke rehabilitation provides its own set of unique challenges when exploring modeling approaches. Historically, rehabilitation data is quite scarce. Rehab patients typically only interact with their caregiver or a clinical researcher once or twice a week for around an hour, and most of the rehabilitation exercises are performed at home. As such, data points collected by the medical professionals are few and far between, and adherence to the rehabilitation regimen is relatively unknown. An in-home device that records user data in real time erases both of these issues related to data scarcity, and allow an algorithm to study individual user’s history to determine what does and does not work for them, as well as study the user base as a whole and provide useful baselines for new users.
In this space, I hope to share with you our journey from small data (data collected from a select few devices) to big data (data collected from thousands of devices). It is through this transition from small to big data that we will be able to truly maximize the benefits of our already impactful rehabilitation devices. Beyond sharing this transition, I hope to provide high-level details about the methods we use to tailor therapy to each individual user and provide meaningful insights from analyzing the rehabilitation data collected by our devices.