The project that we are going to discuss today is basically known as Neurosense, and it’s a multi-year effort started in 2023. As I said, it runs out of COMSATS Islamabad in collaboration with ICESCO, which is a 54-member state organization. And basically, we were aiming at building a lightweight industrial tool that can detect and stage Alzheimer’s disease. And this was, we were doing it by combining three very different modalities...
The project that we are going to discuss today is basically known as Neurosense, and it’s a multi-year effort started in 2023. As I said, it runs out of COMSATS Islamabad in collaboration with ICESCO, which is a 54-member state organization. And basically, we were aiming at building a lightweight industrial tool that can detect and stage Alzheimer’s disease. And this was, we were doing it by combining three very different modalities. And the first one was MRI scans. The second one is acoustic signals. And the third one, which is the hard one, it is the genetics. So the core idea is that no single modality captures the full picture of Alzheimer’s, of the neurodegeneration. So this is what we experienced in the last three years of our project. But together, they let us push, you know, the detection we know much earlier than the traditional clinical diagnosis. So it’s like catching the genetic risk, maybe at birth time, or maybe, you know, subtle speed changes years before the memory symptoms appear. And on the MRI stage, we are having, you know, the structural change in the brain before they become irreversible. So that is basically the idea of whole of what the neuroscience is built upon. So if we go back, we started working in 2023 and that was basically in computer vision. So we use MRI, we develop a multi-class classifier to analyze the structural change of the brain atrophies. So one of our key contributions was here that we adopt the absolute difference marking, which is ADM method. And we utilized it as a post-hoc analysis technique where we subtract a healthy reference scan from a patient scan. So the difference is taking on a pixel-by-pixel level, which produces a heat map. And that’s visually highlights the atrophied brain regions. So it’s, you know, essentially making the models’ reasoning interpretable to any clinical doctor. So we also focused heavily on making the model lightweight enough to run on both edge and the cloud devices as well. So that that was basically in 2023 and that was for MRI but then once we are done with this one we started with the acoustic modeling in 2024. So we extended the diagnostic pipeline to speech so we are trying, you know, functional level features and that happens on across the four domains. So first one was like frequency, intensity, temporal and spectral. So we call it FITS framework. So the findings were very striking that the Alzheimer patients basically show a lower or more variable pitch, which refers to a vocal tremor. And it is according to our statistics that we get from the data in our results. That is about like there is 36.2 percent average decline in the Alzheimer patients compared to the controls. So that was about like on the fifth level. And the second one we are having like on the loudness. So it was, as I said, that on the intensity level, it was 36.2. But when we studied the temporal stability it dropped it to 30.2 percent difference so compared to the healthy controls so it refers to delays in the speech initiation and more pauses, okay. So more pauses mean the persons who are having Alzheimer’s, they take a lot of pauses when they are speaking and why it is happening because maybe as you know that the employed beta or the proteins, they are stuck or integrated on the part of the brain where it is responsible for the memory or the recall of the vocabulary, you know. So, that was basically on the temporal stability. So, and we have measured a very good response in context of acoustic modeling because it is like already very, very cheap because to record a voice you just need a microphone and it is genuinely non-invasive and it is low-cost biomarker that could be captured from something like a normal, a telephone or a smartphone or just like a laptop. So that was we have done on the acoustic model.
Next, we are having genetics model. So that’s refers to basically genomics AI. And this was the hardest piece. We tried three different strategies. First one was on the patient level it was a binary classification and we have you know trained our model and it got to a very low accuracy on a small patient data set that was referring to 228 patients. And the second strategy that we adopted was S&P distribution, zygosity analysis. And in that analysis, basically, again, we are having a very, very low accuracy. If I recall, it was about 60 or 70 percent. But finally, when we researched and investigated about these type of strategies, we got into a genotype prediction. And that was across 77 million SNP data points. So basically, we have this strategy that has been deployed on the Neuroscan and it was the best performing approach and basically we have used basic machine learning methods and I call them old methods not the old methods I call them gold methods basically they are the old methods but I call them gold methods. Why? Because they are not like the generative LLMs because in generative LLMs, you are having a lot of things that is concerning with the responsible AI because they have the ability to generate, you know, across the use case that you are studying. So in that case, you can generate harmfulness and it is also related to the responsible AI. In compared to the old or the gold methods, you do not have that generative capability. So you can use those old methods to train on your specific use case. In our case, it was genetics, it was acoustics, and it was MRI. And we have achieved the accuracy up to 90%. And that is the state-of-the-art benchmark accuracy that we have achieved. So finally, in 2026, whatever we have incorporated from the MRI, from the acoustics, and from the genetics, we have fused all the three modalities into a single tool, which is known as Neurosense. And using the stacking technology, or we call it you know the fusion or the late fusion architecture where MRI audio and genetic probabilities feed into a meta learner and then the final diagnosis and the confidence scores occurs. So basically this was meant that the whole picture is covered by the these three genetic, MRI and acoustic modalities. So basically the future implementation I think, before that what we need to do we need to you know, that in AI we call it co-intelligence okay right now our model is just trained on the recorded data okay so before we even go to the hospitals to the clinics or to the doctors what we are thinking that we introduce this tool to the neurologist to the hospital to the clinical physicians who are working with Alzheimer’s and we want to you know make them test our tool so that we can enhance the outcomes of our tool because right now we do not know that whatever the results that we are getting, what doctors are going to say about that. So that is a missing element right now. And that is basically the limitation. And I think that this is very, very important. Whatever the tool that you are building, first of all, especially for the healthcare, because the human lives are involved here. So we need to be, you know, tested in real clinical environment. And that’s very important. And I think that that should be done with the collaboration with different associations and the Alzheimer International Association thats working with Alzheimer’s.
This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.
