Good day, Monisha Yuvaraj. I'm Arjun. I've had a look through your background and your professional record. I'm interested in hearing more about your work and how you see yourself fitting in here with our faculty. Ready to start?

Hello.

Please start by introducing your professional journey in academia.

I'm Manisha Yavaraj. I'm currently finished my PhD in Biomedical devices and technology from IIT Madras and CMC Vellore. It's a joint PhD program between 3 institute, IIT Madras, MC Valu and Sri Chitra Tanar. My supervisors are Varadhan Eskim from Neuro Mechanics lab, Applied mechanics department from IIT Madras and Doctor Shiv Kumar Balasubramanian.

From the Department of Bioengineering in CMC Vellore.

I mostly worked on using AMD signals to detect movement intention in severely impact stroke patients with movement for assisted and neurorehabilitation. I did my dual MY. I had a dual degree Btech in EC plus Mtech in signal processing and communication system design from triple ITDM Kanchipuram. It's.

I did, it's a five year program from 2014 to 2019 and after that I worked as a.

Project Associate and IIT Madras and 5G testbed after which I joined PhD in 2020 and I defended in November 25.

Yeah, that is most of my academic journey. Currently, I'm working as a postdoctoral fellow in my rehab group, CMC Vellore. I'm continuing on the project.

And.

To look at, I'm currently working on analyzing HDMG signals and that's more about my journey. Thank you.

Thanks for sharing that, Professor Yavaraj.

Professor, let's begin by discussing your research expertise—particularly in artificial intelligence, health informatics, and computer science. Can you elaborate on your specific focus areas and how your work contributes to advancements in these fields?

So as part of my PhD thesis, I worked on detecting movement intention from severely impaired stroke patients I recorded. So generally for these subjects, the residual EMG is going to be really weak, the SNR is really low. So there was.

It is a challenge to detect the intention in such patient using MG signal because first of all the question is whether there is.

Whether do they actually have EMG or not? So that's a question. And we, so I started with different algorithms. So my expertise in machine learning or artificial intelligence was in applying these algorithms to detect their intention. So initially we started with generating these. So synthetic, we use synthetic EMG data you that was generated using Gaussian mod, Gaussian Laplacian and.

Biophysical Model.

And I apply different algorithms which had, uh, both signal processing techniques and some of machine learning techniques. Uh, but because we wanted to implement it in real time, we concentrated mostly on, uh, the algorithms that work. I mean, it was more simpler and does not have much computational complexity. And we found that a threshold based algorithms are much or much suitable for. This application than the machine learning algorithm, but my expertise in artificial intelligence I tried to use.

Gaussian mixture model to detect the, uh, intention in this uh, subject. And also, uh, the next step was to. So in all these case, because it was synthetic data, I had the ground truth information about when the muscle activity was present. But in reality, when we have to apply it for robot assisted therapy in real time, we do not have the ground truth about when the patient is attempting to move because these are all severe patients with no movement. So we wanted to come up with an approach.

That's like an unsupervised approach where even without knowing the ground truth about the presence or absence of muscle activity, we want to detect find out which detector can effectively pick up the low SNR injury. For that purpose I used some statistical distance measures like the.

TBD that is the.

Total variation distance and also other methods like taking the we got. After finding out the binary output, we computed the detection probabilities and using the detection probability I had come up with some measures that can help in identifying the best detector without having the ground rules about the presence or absence of the muscle activity. So apart from that I have done.

A courses apply linear algebra and data analysis. I was the TA for this course for three times and so I have bit knowledge about optimization and linear algebra that is essential for machine learning and artificial intelligence and I'm still continue because it is not.

My major discipline was in signal processing, so I'm continuing to learning machine learning and artificial intelligence. So.

Thank you.

Thank you for elaborating, Professor. Your work on detecting movement intentions in severely impaired stroke patients is fascinating and highlights a critical intersection of AI, signal processing, and healthcare. Would you mind unpacking how you balanced computational efficiency with detection accuracy in threshold-based algorithms versus machine learning techniques during this research? Specifically, what metrics or benchmarks guided your decision-making?

So we for the main objective was to implement the use EMD signal to detect intention and thereby control the robotic assistance. So the robots should provide assistance only when there is an intention to move. So as this detector that I'm talking about which maps the raw MG data to a control signal for the robot should be in such a way that it will as soon as there is an intention to move, the robot should assist and.

Assist only when there is an intention to move. In that case, we want these algorithms to work in such a way that they have a very low latency and high accuracy. For this purpose, we came up with a metric called the detection cost. That is the performance measure that we use which is which depends on three factors one is latency, false positive rate and false negative rate. So the cost measure because the false positive rate and false negative rate are between.

Zero and one we also normalize the latency. So it's a normalized latency by 250 milliseconds because in applications it is said that the person feel in control of the movement on movement up to a delay of 250 milliseconds. So we kept that as a upper threshold. Any delay that is more than 250 milliseconds will be deemed as there is we we as versus 250 milliseconds. So we don't want such high latencies. So all the latencies was normalized between zero and one so.

My cost metric was maximum of latency, false positive rate, false negative rate. So this whichever algorithm that had a lowest detection cost was chosen as the best detection method. So among the 13 detection algorithms that I compared the detector called modified address which does simple rectification and oposs filtering and then computing the threshold from the baseline.

This worked. Had a lower latency, false positive rate and false negative rate. So we.

Consider this as the best approach for robotic therapy and also about the computation efficiency when we looked at so the number of parameters that needs to be optimized for this approach.

Is also lesser, so while training, the number of computations that was taken was much lesser than the other. Machine learning algorithms that was used. So even for real time application the computation efficiency.

Of this, this method was better.

Understood, Professor. Your focus on optimization—especially in balancing latency, accuracy, and computational efficiency—is pivotal for real-time applications like robot-assisted therapy. The detection cost metric you developed is an elegant solution to integrate multiple performance factors. Moving forward, can you elaborate on how your expertise in statistical measures like total variation distance and unsupervised approaches contributes to expanding the applicability of such technologies, especially when ground truth is unavailable?

This is a important issue that is there in the current mesh learning algorithms.

But I would not be very because my application was not. I did not really use the actual supervised methods in the literature I it was only I can call it unsurprised just because I did not use the counter otherwise it was not majorly the traditional algorithms that is used.

Thank

So I would not.

Claim that I I am well versed in this machine learning algorithms.

About the question of how the total variation distance or the so the total variation distance was used to check how different is my rest state AMD data and move state the EMG data. So what I mean by unsupervised method here is that in all these cases we recorded data from the subjects when they were asked. Like in each trial the subjects were asked to relax for some time and move for sometimes.

So we had the rest period and.

I understand, Professor. You utilized the total variation distance as a statistical tool to differentiate between rest and movement state EMG data in your research. This approach provides valuable insights, especially when ground truth labels are unavailable or unreliable. How did this method perform when applied in real-time robotic assistance scenarios, and were there challenges in ensuring consistency across different subjects?

So I did not apply this algorithm yet on the to control the real time robot yet. But on computing the detection cost for the algorithm when compared to the simple threshold based method, total variation distance had a higher latency. So we did not carry forward the total variation distance method to apply for robotic therapy. As I said before, I used only the threshold based detector. For controlling the.

Rob So, uh, controlling. So I did not implement the robot control yet. I just tried to control a cursor of a game using this algorithm and that uh, even with modified artists, because the uh, EMD signals was highly variable, the control is still difficult and that is quite challenging.

And I'm still working on coming up with a better way of controlling the report.

Got it, Professor. Your work on cursor control using EMG signals is an important interim step before applying the methods to robotic therapy, and it clearly highlights the inherent challenges of variability in EMG signals. Given your ongoing efforts, how do you plan to address the variability and improve the robustness of your detection algorithms for future applications in robotic control?

So one way is so there is always a trade off between the the.

Smooth so the variability and latency. So if if we if the signal is too variable, we'll have to apply.

Higher filter filtering techniques that will of course increase the latency. So this trade off is always there. So one approach that I thought would be useful is to get the patient feedback itself. Like once when you apply the robotic control with a certain parameter, let it.

We asked the patient for their feedback whether they feel in control of their own movement. If not, we adjust, we put in some time, filter in, in, in between with some window size and we have this window size variable so such that we keep changing the window sizes and ask feedback from the subject itself. The point where the subject says, yes, I feel in control or I have lost in control of my own movement. That is end of the day, it's all about how natural. Naturalistic The human machine interaction is so we.

Even though if it is if it has a higher variability.

Take into account or. We keep the patient in the loop and then decide, uh, whether that particular algorithm works or not. So that's the one approach that I thought would be useful for this purpose.

Thank you, Professor. Incorporating patient feedback directly into the loop to optimize human-machine interaction is indeed a thoughtful and patient-centric approach. Balancing variability, latency, and the naturalistic feel of such systems is critical for advancing therapeutic applications. Let me now shift focus slightly—earlier you mentioned your teaching experiences as a TA in linear algebra and data analysis. Could you describe your approach to teaching complex topics like optimization or matrix mechanics to ensure students grasp the underlying principles effectively?

Uh, so always the fundamentals is important. So I would start, uh, with basically linear algebra and then, uh, at least for the first few classes and then move on to the optimization approach. And one thing that as RTA and as a student that I learned is practical way. So, uh, my professor generally have a lot of programming assignments and then illustration using examples. So I thought.

I would also follow the same way of teaching optimization with the graphical.

I mean something that graphically explains how optimization works that I think help can help students learn better and most assignments or the tutorials with which are application oriented can help students understand better. So I think that approach I would opt for.

Understood, Professor. An emphasis on fundamentals, graphical illustrations, and application-oriented tutorials is indeed a structured and effective approach to teaching complex concepts like optimization. It helps bridge theory with practice, ensuring deeper understanding. Could you provide an example where you employed this method during your TA experience, and how students responded to it?

Like not only some of tutorial sessions and uh, I have solved basic uh, questions like, uh, but one in one of my other class that I took on single processing, I explained what, how an exponential signal or a complex exponential signal look like. This is something that I did not learn during my, uh, undergrad. So, uh, when I could explain how the exponential signal actually looks like, it clears a lot of configurations about.

So I had taken.

How it actually present. So when I had to plot the exponential signal in Matlab and showed it to the audience, I think they understood better on how actually the complex e-mail looks like and that that is one example of how graphical plot can explain in. During my Aladdin tutorials I had only solved the problems on the board. I have not.

Done any graphical illustration from my my.

Thank you, Professor. The use of MATLAB to visually illustrate concepts like complex exponential signals demonstrates your commitment to enhancing student understanding through practical and graphical explanations, which can significantly help in unraveling abstract ideas. Now, shifting gears, one of the critical aspects of this role is guiding student projects and fostering impactful research. Could you detail an instance where you mentored or guided student projects, and how you ensured they developed both technical expertise and critical thinking?

Uh, during my PhDs I I.

Did not have much experience about guiding anyone but I I had helped a few of my master students for traveling PMG preamplifier board and also for using EEG signals to I mean electrodes for EEG signals. This is something I have helped some of the master students and currently I I'm helping one of my PhD juniors to.

Do I mean do the HTMG recording for mirror therapy?

Uh, experiment. So, uh, these are the few things, I mean, uh, some of the examples that I have guided my, uh, fellow mates. And apart from that, uh, I have, uh, conducted, uh, Jana club activities in my lab where we read papers. And in that cases, I have helped my fellow mates of how to read a journal, how to present it and how to. Uh, summarize a general paper and I've also helped some of my.

Fellow mates in writing some of the journal articles and other grants. Sorry other.

How to search current etcetera so.

That's most important.

Thank you, Professor. It's clear you have played a supportive role in guiding peers and juniors in experimental setups, journal club activities, and writing processes. These experiences align well with the expectations of mentoring students and fostering collaborative research. To probe further, given your involvement in journal clubs and your support in writing articles, how do you approach teaching students the ability to critically evaluate research papers effectively?

Uh.

I see you're completing your thought—please continue.

In general, to start with reading as article is always for me at least, it was very difficult in the initial stage. So especially when you have to read a lot initially. In my initial days I used to take mostly when I tell my fellow mates, it's mostly about my experience of how I overcame and how I improved reading journal papers. So one approach, initial days I used to spend like half an hour or the whole half day.

Yeah. So for the.

For reading a paper, but that that doesn't seems to work.

One approach is to skim read the paper. That is you first read the abstract. If you find it in this thing, go read the result and try to connect the figures in the result. That is what I generally suggest. And then if you find then while reading the result itself, you have several question of how how this was done and then you go connect the method section. And it is also one way is also to read the last paragraph of the interaction where they give the scope of the paper that is.

So trick to quickly read the paper. Yeah, this is most of how I tell my fellow mates of how to read a general article quickly and whatever we learn if it if it is in case of if it is for your work, I generally recommend people to make a excel sheet to summarize this whatever they have learned from that paper that is what is the objective of their own work and how this paper contrib.

Contradicts their idea how it is, how your work is going to be normal from that paper. Is there something? These are the points that I I always recommend my fellow mates to note on SO.

Thank you, Professor. Your structured approach to reading research papers—starting with key sections like the abstract, results, and discussion, and connecting them to methods—is a practical strategy that certainly helps in extracting relevant insights efficiently. Encouraging the use of an Excel sheet for synthesizing key points and contrasting ideas further reflects your commitment to developing critical analysis and organizational skills among students. Let me now shift focus to your expertise in research publications. Could you share an example of a paper you authored in a reputed journal, focusing on its impact or contribution to the field?

I have two general publications now and one conference from my PhD and one conference from my masters. 1 international conference from PhD and one national conference from my masters SO.

The two journals that I had public published in, one was in F-1000 research, it is a human journal, and then the other one is I Triple Access which is also a human Scopus index Q1 journal.

Uh, and I think in the both contributes to the field of uh, so there isn't much research that has explored therapies for severely impaired patient with no movement. And both papers focuses on uh, using AMD signals for this purpose. And uh, because I have shown that there can be methods that can pick up residual EMD from severe patients, I think that contributes to the field like people can explore.

Other therapeutic interventions for severe patients.

That is 1 starting point that a simpler method like AMG that is suitable for routine clinical use can be used for.

Robotic therapy for severely impaired stroke patients. So I think this is the starting point. This will be a starting point for people to further explore therapeutic intervention for severe patients.

Thank you, Professor. Your research on utilizing EMG signals as a foundational method for robotic therapy in severely impaired stroke patients is indeed impactful. By demonstrating the feasibility of these techniques for routine clinical applications, your work opens the door for advancing therapeutic interventions and enriching the field. Building on this, can you share insights into your experiences with securing funding grants or collaborating with industry partners to expand the scope of your research?

I have not much collaboration with the industry partner, but I have collaboration with some of my professors, previous professors, post Doc guide and his colleagues. So I had traveled to Imperial College London because I had a travel grant for one month and I had the opportunity to work with Professor Ithi and Burton from Imperial College London and I also traveled to. University in Sweden, where I had worked. I mean, I was there for a week.

Again on a travel card. Yeah. And then I have the PMR fellowship, Prime Minister Research Fellowship. That is one thing that I'm that I had got during my PhD.

Thank you, Professor. Your collaborations with renowned institutions like Imperial College London and universities in Sweden, supported by travel grants and the Prime Minister Research Fellowship, highlight the international scope and recognition of your research endeavors. These experiences undoubtedly enhance the depth and reach of your contributions. Let us now discuss adaptability in teaching laboratory courses, a crucial component of this position. How do you approach ensuring that students gain hands-on proficiency while maintaining a strong connection to theoretical concepts?

Uh, more of, uh, so, uh, especially in this field of artificial intelligence, I think more of, uh, exposure to, uh, data analysis is a must. So, uh, I think code most practical way is to get exposed to mute the existing state-of-the-art algorithms and to start understanding them first try them out practically.

And a decent coding skills is something that is required, so I think I would focus on.

UH developing UH students coding skills and then exports it to UH recent state-of-the-art algorithms.

Thank you, Professor. Your focus on building coding proficiency and exposing students to state-of-the-art algorithms reflects a practical and forward-looking approach for laboratory courses, especially in fields like artificial intelligence and data analysis. Let me pivot slightly—effective communication is critical in academia. How do you ensure clarity and engagement during lectures, particularly when explaining complex concepts?

Uh, I have not much, uh, I don't have much experience with teaching except for some of the tutorials that I have taken for the students as a TA. Uh, so I think that I would learn over experience from my, uh, whenever, when I get a position, uh, apart from that, a good presentation has skills is really required to deliver whatever we want to.

Convey to the audience. So I always concentrate much on my presentations.

Whenever I do a one, uh, there, I take a lot of effort in uh, graphically or pictorically explaining concept to the people whenever I present. So I prefer doing that over having a lot of more contents in the presentation. I think a good presentation skills is essential for conveying whatever you want to communicate to the audience. That is something that I would concentrate and I think this is something that I have not experienced so far so.

So I would like to, I mean like to learn it over time when I get a position.

Thank you, Professor. Your emphasis on strong visual presentation skills and a clear, pictorial approach to simplifying complex concepts is indeed valuable for effective communication in academic settings. It's also insightful that you are open to continuous learning and refining your teaching methods over time. Let me now move to our concluding area—industry collaborations or consultancy. Can you discuss any experiences or ideas you might have for bridging academic research with industry practices to ensure real-world impact?

Uh, I.

So most of the work that I have done is on stroke application. So I think developing.

I mean medical devices or therapeutical interventions for.

Stroke rehab especially my field is on brain computer interface. So I want to have systems that has simplified brain computer interfaces that can go into the clinics and the markets. And also I'm interested in developing better EMV sensors, wireless ways that can be used for regular monitoring at home and at clinic. These are some of the things that I think.

Can help me collaborate with industry and with the hospital settings?

Yeah, that's most important.