My Postdoctoral Research
After finishing my graduate work, I was very fortunate to receive a postdoctoral fellowship from the Translational Research Institute for Space Health. This gave me the wonderful opportunity to sow and germinate the seeds of ideas that I incubated in my graduate work and guide its growth into translational research that addressed core risks and gaps outlined by NASA's Human Research Program. To understand this section, it will help to either read about my graduate work or watch the video in the Introduction.
Project 2.1: Using Machine Learning to Characterize and Predict Individual Differences
In my graduate work, I strapped blindfolded participants into a machine that was programmed to behave like an inverted pendulum. The participants were instructed to use an attached joystick to stabilize themselves around the balance point. In my spaceflight analog condition, they were placed in the Horizontal Roll Plane where they no longer tilted relative to the gravitational vertical and therefore could not use gravitational cues to determine their angular position from the balance point. In this condition, they became spatially disoriented and crashed the machine often (crash limits were set to +/- 60 deg from the balance point and when participants reached that limit, the machine stopped and reset back to the balance point). Several participants reported unusual illusions and everyone showed a characteristic pattern of positional drifting. Collectively, they showed very minimal learning even after 2 experimental sessions on consecutive days. However, when I looked at the individual data, I was surprised to find a wide range of performance abilities. I could tell there were secrets hidden deep within the individual differences beyond the general clouds of averages and standard deviation.
There was a large range of individual differences in the spaceflight analog condition
Each circle is a different participant that I ran in my experiment. On the x-axis, the "90deg Roll" refers to my spaceflight condition (Horizontal Roll Plane) and "0deg Roll" refers to the control condition (Vertical Roll Plane) where participants could use gravitational cues and had a relatively good sense of where they were. If a circle is above the red line, it means that they learned to reduce the variability of angular position. If a circle is below the line, it means that they actually got worse and the variability increased. In the control group, everyone learned, however in the spaceflight analog condition, some people got worse whereas some people got better. It should be noted that the best learners in the spaceflight condition still performed much worse than the control group. To reiterate, when I only looked at the group averages, I concluded that there was very minimal learning in our spaceflight analog condition, however the individual differences revealed a much deeper story. Somehow a few participants were able to learn and improve in this harsh disorientating condition and oddly others became much worse even with time to practice.
We used machine learning to cluster participants into 3 performance groups
Once I had identified that there was a wide range of abilities, my next goal was to determine whether people could be clustered into different performance groups. This was a difficult task because I had more than 20 metrics that quantified performance. In partnership with Pengyu Hong (a Brandeis CS professor) we wrote and obtained a small synergy grant from TRISH. Our first task was to use machine learning techniques to cluster people into 3 statistically distinct groups: Proficient, Somewhat-Proficient and Not-Proficient (you can get the details in the paper at the end of the section). To verify whether the clustering has meaning, the graph on the right, shows two important metrics: on the x-axis is variability of angular position and on the y-axis is the frequency of crashes. You can see that those who are in the Not Proficient group (squares) have large positional variability and number of crashes, whereas those who are in the Proficient group (stars) have small variability and crashes.
We used machine learning to predict each individual's final performance
Once we had clustered individuals into 3 statistically distinct performance groups, our next question was, could we predict each participant's group (Proficient, Somewhat-Proficient, or Not-Proficient) at the end of the second day of experimentation? While this is a standard machine learning question, it was difficult to do with our data set. In most machine learning implementations, you have 'ground truths' that are used to train the classifier. In our case, it was not always obvious which group a person belonged to. Another problem was that our data set was not huge. You can read our paper (below) to see how we addressed these issues. We found that we could predict a person's final performance group with greater than 80% accuracy using data from the first 10 minutes of the experiment.
Those who were Not-Proficient were not randomly bad, instead, they all used the same suboptimal strategy!
Participants who were in the Proficient group (top graph) showed learning across majority of metrics, which was surprising because most reported feeling disoriented in our spaceflight analog condition. It should be noted that even though the Proficient group showed learning, their performance, compared to the control group (Vertical Roll Plane), was still much worse. The Proficient group, had relatively small oscillations of angular position (black line) and they used small varied joystick deflections (grey).
Much more surprising were the findings with the Not-Proficient group (bottom graph) who, over time, became worse on almost every metric....except for one metric. They all learned to decrease the number of crashes! This was so unbelievable to me because they were somehow crashing less but they were making larger and crazier movements with the machine (black line). I call this a suboptimal strategy because they improved just one metric at the cost of making everything else worse. What was even more interesting was that everyone in this group was using the same strategy: they smashed the joystick (grey line) back and forth as hard as possible in a very stereotyped manner. This led to the question, could a training program push these participants out of their suboptimal strategy? Find out in the next section. Also, check out the paper for the details on how I quantified stereotyped large magnitude joystick deflections using the Hilbert Transform.
paper5 rough draft.pdfProject 2.2: Developing a training program to enhance performance in our spaceflight analog task.
The machine learning project had revealed that there were huge individual differences in learning and performance of our spaceflight analog task. Participants in the Proficient group actually showed significant learning. This was surprising because 90% of participants reported spatial disorientation and did not having a good sense of where they were. In contrast, the Not Proficient group became worse on all metrics except one. They all learned to reduce the number of crashes at the cost of everything else. They did this by using the suboptimal strategy of making very large stereotyped joystick deflections.
This led to the next question: Could we develop a training program where we simply told the participants not to smash the joystick? To answer this, I ran an experiment where I gave participants several strategies that were collected from the Proficient group. These strategies included advice on how to use the joystick along with ideas distilled from my previous work on the 2 dissociable components of balance control (read the paper below for more details). However, knowledge of the correct strategies had no statistical effect on learning and performance.
Why were people not using these strategies? In most motor learning studies, you give a person a strategy and they try to implement it. As they make errors, they use the strategy and correct themselves and eventually learn the task and reduce their errors. In our experiment because participants are blindfolded and because they cannot use gravitational cues, they don't have continuous feedback on how well they are doing the task...until it is too late and they crash the machine. That is, it is possible that they are unable to apply the correct strategies because they don't have the feedback required to update their internal models. Using this insight, could we create a training program where participants have access to continuous feedback of their position? And could we provide that feedback in a condition similar to the spaceflight analog condition where participants cannot use gravitational cues to do the task?
We developed a training program where participants received continuous feedback on their angular position from gravitational cues however could not use that information to do the balancing task similar to the spaceflight analog task.
We did this by placing people in the Vertical Roll Plane, where they can use gravitational cues to obtain a good sense of their angular position. In the past, when we ran experiments in the Vertical Roll Plane, we always had the balance point at the gravitational vertical. However, to simulate a condition similar to the spaceflight analog condition in the Horizontal Roll Plane where people can only use motion cues to find the balance point, we randomized the location of the balance point. For example, in the figure to the right, the balance point is where the red line is. Participants were never told the location of the balance point which changed every trial, so they had to search the entire space, focusing on motion cues (such as detecting when the direction of falling changed, indicating that they just passed the balance point).
Those that received this training performed much better in the spaceflight analog condition when compared to those who did not receive the training.
In the table to the right, the control group remained in the spaceflight analog condition for the entire experiment. They were given strategies before the experiment began. The Vestibular group were place in the spaceflight analog condition in the beginning and then again at the end. In between they were given the training program described above. You can see that they showed significant improvements across many metrics. This means that our training program was effective.
There was long term retention of the skill acquired in the training program
Because I am funded by TRISH, my research is focused on applications to spaceflight. The previous results show that my training program helped enhance performance in a spaceflight analog condition where participants become spatially disoriented similar to what astronauts may experience when landing upon the Moon or Mars. However an important question is how long do the effects of the training last? For example, it will take astronauts several months to reach Mars, would the benefits of the training last that long or will they have to be retrained?
Surprisingly, there is very little literature on long term retention of skilled motor learning. Additionally, my task is different than those in the prior literature because the condition where the skill is learned is very unique and not typically experienced in normal life. To study this, we brought participants back after 4 months and put them into the spaceflight analog task. We were surprised to find no deterioration in their abilities which means that they retained the skill across a long period of time.
paper5.pdfProject 2.3: The Role of Spatial Acuity in Individual Differences for Balance Performance in our Spaceflight Analog Task.
I provide a quick overview of the project and will provide more details after the paper that I have submitted is reviewed and published.
In Project 2.1 I shared that participants in the Not-Proficient group acquired a suboptimal strategy that helped them reduce the number of crashes at the cost of everything else. They would smash the joystick back and forth which would cause the machine to make huge amplitude oscillations. The first time I observed this behavior, I thought the machine would break because the oscillations were so intense. Why would all of the participants in the Not-Proficient group converge to this strategy?
One idea is that their vestibular thresholds are much higher than other participants. That is, they may need much higher velocities to accurately detect motion and that might be why they enter into large velocity oscillations. Similarly, it might be that these individuals have a poor sense of their spatial orientation especially in our spaceflight analog condition where participants are highly likely to become spatially disoriented.
How could we measure a participant's spatial acuity (i.e. how accurately they can determine where they are in space)? To test this we designed an experiment where participants were blindfolded and placed into the same machine. Instead of controlling the machine using a joystick, they now just passively sat in it and experienced different motion profiles. Their only task was to press a trigger button every time they felt like they passed the start point. Very surprisingly, we found no correlation between a person's spatial acuity and their ability to actively balance the machine in the spaceflight analog condition. This most likely means that a complex balancing task in a disorienting condition depends on many factors, and an individual's limitations in spatial acuity does not solely predict their ability to learn and perform our spaceflight analog task.
As I dug through the data, I found some unexpected correlations between spatial acuity in the Vertical Roll Plane where people have gravitational cues (and a good sense of where they are) and balancing ability in the spaceflight analog condition (Horizontal Roll Plane).....but only in the later trials. This finding was really unusual because I had never previously thought that balancing in the Vertical Roll Plane (where you have gravitational cues) could have relevance for balancing in the Horizontal Roll Plane (where you do not have gravitational cues). Additionally, because the correlation was only significant for later trials, it suggested that perhaps fatigue was playing an important role.
To test the role of fatigue, we ran another experiment. First, the blindfolded participants did the passive task in the Vertical Roll Plane where they have gravitational cues and a good sense of where they are. As a reminder, their only job was to press a button every time they felt as if they were at the start point. Then we put them into a different chair (pictured to the right) that spun in Earth Vertical Yaw as they moved their head up and down once. This induced a sense of early motion sickness and fatigue. Then we brought them back into the original machine and had them do the passive trigger pressing task again. We found significant correlations between a person's spatial acuity in the passive task (Vertical Roll Plane) and their ability to perform the balancing task in the spaceflight analog task (Horizontal Roll Plane). This suggests that attentional or broader cognitive resource capacity is very important when initially experiencing a novel disorienting situation.
