173: AI and the Human Touch: Patient Safety, Prognosis & Voice Biomarkers

How far can AI go in helping us diagnose disease—without losing the human judgment patients rely on?

In this episode, I break down four studies shaping the future of digital pathology, oncology, and neurology. From spatial biology updates at SITC to voice-based Alzheimer’s detection, deep learning for sarcoma prognosis, and new guidelines for safe AI deployment, this week’s digest highlights where AI is making a real impact—and where caution still matters.

Episode Highlights

1️⃣ SITC Trends & Spatial Biology (00:00 → 07:40)

I share key updates from SITC 2025, including the growing role of multiplex immunofluorescence (mIF) and the need for integrated staining-to-scanning workflows. I also preview new educational content and upcoming podcast guests in global AI research.

2️⃣ Digital Neuropathology & Alzheimer’s (07:40 → 13:01)

A major review confirms that digital neuropathology is now robust enough for large-scale Alzheimer’s studies—opening doors for computational tools to link histology with cognition.

3️⃣ Patient Safety in AI (13:01 → 19:56)

An Italian review underscores the foundations of trustworthy AI: dataset quality, transparency, oversight, and continuous validation. I discuss why “patient-centered AI” must remain our standard.

4️⃣ Voice Biomarkers for Cognitive Decline (19:56 → 26:43)

AI models analyzing short speech recordings are showing high accuracy for early Alzheimer’s detection. This could make future screening simple, noninvasive, and more accessible.

5️⃣ Deep Learning for Sarcoma Prognosis (34:06 → 35:59)

A multi-instance CNN outperforms FNCLCC grading by identifying prognostic patterns in tumor center and periphery regions, offering new insights into soft-tissue sarcoma biology.

Takeaways

• mIF is maturing quickly but needs standardized, end-to-end workflows.
• Digital neuropathology is ready for broader Alzheimer’s research.
• Safe AI requires multidisciplinary collaboration and rigorous validation.
• Voice biomarkers may become powerful tools for early cognitive assessment.
• Deep learning can refine prognosis and reveal hidden tumor patterns.

Resources

Hamamatsu (MoxiePlex) • Biocare Medical (ONCORE Pro X) • SITC Programs • Recent publications on AI biomarkers and computational pathology.

Thanks for listening—and for being part of this growing digital pathology community.

Transcript

00:00:00
Aleks: Good morning, trailblazers at 6:00 a.m. from Pennsylvania. I hope you are ready for today's dig digest. As you are joining, I'm going to give you a few updates as always. So, I just came back from SITC just means last week, last um last Friday, Saturday, it was SITC annual meeting, Society of Immuno Therapy and Cancer. And let me switch on the chat because ah I already have people sending me comments. Thank you so much. This is amazing. Uh let me know that you're here. Let me know where you're tuning in


00:00:41
from, what time it is, where you are. Uh I can see the comments definitely from YouTube. Um probably LinkedIn. A few of you are there as well. So yeah, SITC Society of Immuno Therapy and Cancer. So um how I see this like area of spatial biology uh spatial biology, immunofluorescent multiplex immunofluorescence and is it kind of informs the um cancer research cancer treatment and uh future clinical tests although okay we have guests from Germany amazing um future clinical workflows which like traditionally were


00:01:30
then being translated into either uh IHC or yeah IHC most of most of the time but now the complexity increases so they're I think they're going to be uh just native clinical uh immunofllororesencebased uh tests and we already have you know we work with fish we work with whatever I think fish is the the the most common fluorescent thing that is happening uh in the diagnostic space but there's going to be more because we learn more biology uh the medicine is more personalized so that's what sits was


00:02:11
about let me just adjust this camera so that it follows my face okay um so yeah that's where the the the field is heading and uh So I was there with Hamamatsu and with Bioare. Hamamatsu has this multiplex uh scanner that scans up to 10 markers. It's called Moxiplex. And um what they did they had this they have this partnership with Bioare which is a staining company. they have a stainer and they're going to be putting together um not only them but in general that is a trend putting together


00:02:50
end to end workflows for these more complex uh tests that are based on immunofluorescence, right that's uh a little bit more complex than brightfield so you do want to have your partners lined up for each step of the workflow and one thing that has nothing to do with immunoncology that I learned at SITC is that you can rent plants for the conference booth. Did you know that? So, Hamamatu had this like bromeliad plant and I was preparing for a live stream and I'm like, "Ah, let me put this plant somewhere." But I'm like,


00:03:25
"Oh, how come they have a plant? Did they bring it from home?" No, you can rent it at the conference. I thought that was fantastic. Like, people are brilliant with their business ideas. So you can decorate your booth with plants and you don't have you can just rent them there. Um also another thing that you're going to be getting if you are on my email list and if you are not I'm going to give you a chance to do that right now. Uh wrong code because you can get on the list when you get the book. So there is


00:04:00
a book digital pathology 101 that I'm actively working on updating but uh if you get this book through the code you're going to be on my email list and the email list is going to get uh email list gets all kinds of cool stuff that uh I come up with but uh specifically from sity you're going to get a recording of the talk Dr. Carlo Befulkco gave a talk uh where he also thinks and that's his work where he thinks that there's going to be um with the level of personalization of medicine we are going


00:04:35
to go into immunofluorescent as a diagnostic tool, and he did research on that in his lab and this talk is going to be distributed only to people who are on the email list. So get on the email list if you're not there yet. If you got emails about this live stream, it means you're there. Uh, and you probably have the book, but if you don't have it yet, um, the code is here and I'm going to, um, I'm going to include the code at the end as well for those who join a little later.


00:05:06
Another uh, thing that I wanted to tell you. Oh, upcoming guests. So, we already talked last time that Andrew Janoik is going to be a guest on the podcast. This is already recorded. Then we have one uh podcast coming up with Victor Anhel from OMA.ai. So this company is creating, um is collecting data from a Latin American population in a responsible uh patient centric way to empower AI develop development of AI tools that are u diverse that that um represent populations that are underrepresented


00:05:48
and his story is super powerful how he came up with this. He also has a personal story to tell. So, uh that's going to be coming up. And today I'm recording with the representatives of Big Picture. Big Picture is a a private public consortium in the US uh aiming at creating a huge slide repository like millions of slides for AI model development. And and I'm going to be meeting with Yerun Fanderlac and Julie Baklair who are leading this thing. uh and by this thing I mean this multi-year


00:06:21
consortium that had to um transform and pivot a little bit as they were um as like transformers generative AI uh was being um introduced into the space. So these three uh are coming up and yeah book again let me give you the code one more time. Um, I am committing to updating it uh before the end of the year. Um, actually I have a date. Uh, keep me accountable to this date. On December 1, I'm going to like rent a co-working space. Uh, I'm going to go to Frederick. I'm going to rent a


00:07:04
co-working space and I'm going to work on this book. Um, I already did some work with you online as well. So, I think one to two days should be enough and you're going to be getting the updated version. Whoever already has this version, uh, which the PDF you can download for free through the QR code is going to get the updated version with new AI information and like the most up-to-date stuff in the digital pathology space. And now, thank you again for joining me. And let's go to our papers cuz I have a few


00:07:40
prepared for you today. And we're going to start with uh a couple of of these are um reviews. So, you know, in the review abstracts, there's not so much substance, but what did I do? No, nothing. I thought I destroyed my camera. But um I still want to mention them because they kind of indicate trends. So for example this one uh digital neuropathology of neurodegenerative disorders foundation research advances and future directions. So this is something this theme topic has been coming up at least like I've


00:08:22
been seeing it. Maybe it's the um what is it called? Reticular activating system. when you start working on something then you see it all around. So for example if you buy a yellow car suddenly yellow jeep you see all the yellow jeeps all around and but uh because um I gave a talk on digital neuropathology in the pre-clinical space now suddenly digital neuropathology papers are showing up uh after the conference. One of them is this one from I forgot the water today from the Emory univers University Davis California and


00:09:03
basically US groups and what they say here is that neuropathology examination after death remains the gold standard for differentiating between Alzheimer and um Alzheimer disease and um Alzheimer disease related dementias like from the diagnostic standpoint if your gold standard is after death to differentiate something from something else um not really helpful so um but that's what it is right um and um I have another paper on how to do it actually in life uh later which is super cool because it


00:09:44
talks about voice biomarkers for Alzheimer disease but in this one This review provides over view of digital pathology technologies and their associated infrastructure which uh this is the important thing here because um the the questions or the concerns that I've been hearing about neuropathology was like oh we're not ready brain is too complex brain is uh like we miss something um but the papers that are coming out from these neuropathology group consistently tell us hey we are there already it's good enough uh we can


00:10:20
use digital pathology for neuro and neuroscience has been doing this for years right but obviously um research and diagnostics are kind of different from their regulatory perspective so and I um I totally uh understand their concerns but here in this review uh they talk about the digital pathology technologies uh available open-source and proprietary digital top pathology software relevant background of neuro neurodeenerative hystopathological features. So I think this is cool also uh that they mention


00:10:57
these uh hisystopathological features and computational research. They discuss evidence supporting how recently developed technologies and methodologies can enhance our understanding of histopathological features of neurodeeneration and correlations of hystopathologic features with cognitive performance and age and death. So this is also I will have a bunch of stars but um the correlation of hisystopathologic features with cognitive performance and age at death. So this is something where once you have like the the pairs of data


00:11:37
clinical data and then hisystopathology data and if you have it like close to this time when they are performing the neuropathology examination after that then you may predict one or the other and here rather like predict from the clinical predict the hisystopathology hisystopathology uh that's kind of like telling you Okay. What's going on there? And we're going to talk about it in a second as well. So highlights historical summary of digital pathology with respect to neuropathology. So all my neuro people


00:12:20
have a look at this one. uh key digital pathology technologies that are supporting neuropathology uh and digital pathology applications in neuro degenerative disease um and also discussing the future of digital neuropathology. So definitely something to look at especially if you are active in the neuropathology space as a researcher as a um as a diagnostician as a vendor uh whoever is in there. The next paper that we have is patient safety in AI powered diagnostic pathology. Patient safety uh compliance,


00:13:01
privacy, these are the things that come up and I I had my chat not open and you guys are uh are joining. I have um we have a couple of questions. I'm going to address these questions um at the end so that I can give them the necessary time. I don't know why my font for this is so um like so big. I'll figure it out. Uh patient safety, right? Patient safety. Um GDPR compliance in Europe, HIPPA compliance here in the US. Um and this is a group from Italy. I like when European groups uh are discussing this


00:13:45
um because they are Europeentric. I don't want to like I don't know how to convey that but um often it's it's a kind of more holistic perspective because uh they have this very strong presence of the American US perspective and then they have their own whereas in the US it's like European perspective is less relevant uh for the people here but the ideas and the ways they approach it are uh super important as to incorporate. So let's see what this Italian group did here or what they're


00:14:27
describing. AI powered diagnostic pathology involves combining traditional hisystopiological techniques with computerass assisted AI technology. Uh and there are a couple of steps in this process. We have generating host digital images which is usually the first step that everything is based on. Then we annotate and train algorithms. uh sometimes you need more annotations, sometimes less. Now with the foundation models, the tendency is to go towards less annotations, which is fantastic because you're avoiding a bottleneck.


00:15:01
and then constructing robust data sets. Um that was a topic that was covered at the um American College of Veterary sorry ACVP American College of Veterary Pathologist annual meeting where we had a separate talk on data sets like how important are data sets and giving people access or like making it possible to access data sets to actually create these things and by these things I mean AI algorithms uh AI models digital pathology tools uh create and test them. So that's also big picture consortium


00:15:36
that's going to be addressing this but data sets like the the keyword data sets um I think before listening to that talk it was like in the back of my mind and that talk by Christoph Bertram brought it to the front like how important it is to benchmark your things um and now on a regular basis as well then obviously testing and monitoring consistency and that's what data sets are going to be uh important for um sorry uh testing and monitoring cons consistency with clinical expectations and validating results externally and


00:16:16
overseeing the output of algorithms. Um these steps need to adhere to quality standards and obviously ensure patient safety. Um, and current evidence suggests, and this is a theme coming up over and over again in different papers, is that AI can enhance the accuracy of human diagnostics, enhance the workflow, basically uh kind of like supercharge you as a researcher, as diagnostician, but it cannot replace humans as autonomous classifiers. There are attempts uh I don't trust these attempts yet. I'm


00:16:55
not going to like mention what kind of attempts. Uh I'm just going to tell you that uh I've recently heard about something in the venerary space. Once the publication's out, I'm going to tell you more about it and let's see what they put in that publication. Anyway, so now we have a generative intelligence uh that is promising us technological advancements and uh when applying all these technologies we need to uh adhere to international healthcare institutions recommendations um and they


00:17:31
recommend clearly defining the application domains and implementing and monitoring safety measures. So keyword application domain right it's not going to be one tool that solves everything it's going to be something for a specific application for intended use for a use case like all these words are going to be um there that's what it's going to be for and not like oh one tool that solves all problems h I don't think any of you like believes that we will have one tool that solves all problems


00:18:05
even though chart GPD is pretty powerful Um and we need to focus on patient centered safety considerations. H and that requires the um necessity of collaborative efforts among governments, academic institutions, international healthcare agencies, scientific societies, patient associations, algorithm developers. So like a bunch of different stakeholders. And um when the podcast with Andrew Janoik goes out, you're going to hear more about that. Like when they validated and deployed a digital pathology tool, all these people


00:18:45
in one form or another were involved in um in working on deployment of the tool. On the creation of the tool, it was mostly algorithm developers and the pathologists that were helping to develop the tool. But then deployment, it was like representatives of a lot of departments, right? And um one important thing that he said uh Andrew said during this uh podcast was that not a single person that is on the author list and there's a long author list can reproduce this work that they did on their own. And that is the trend that


00:19:21
I'm seeing as well. That's um something I was talking about at the ACVP meeting as well that to ensure liable and reliable AI uh in whichever form uh or shape you are trying to build it. It's going to not just be like the computer scientist. It's not just going to be the pathologist. It's going to be everyone working together. Uh and uh um that's how it's going to be. especially um legal also it like people that you didn't think you will ever be working with uh based on your specialty they're


00:19:56
going to be your partnersh in in in this endeavor of um providing reliable and liable AI let's see oh this one is so cool guys I love this voice prints of cognitive impairments so that goes together with the first one that we had that well we have to do histop pathology postmorton, but actually we can listen to people while they are still alive and help with diagnosis. So before we dive into this one, um I want to show you something. So there's a tool u that I have this this thing for recording like voice notes and um


00:20:35
sometimes meetings. I mean I have an AI recording my meetings most of the time anyway. But this is more like okay uh I was using at the conference at sity uh we were having different discussions with uh potential podcast guests, potential partners and in general like brainstorming and then um I I talked to it right so this is something that could potentially give me diagnostic insights if I was um some kind of early onset Alzheimer's which I don't think no there's nothing indicating that um


00:21:10
But the point I'm trying to make uh phone is listening to you. So like we have these tools that are basically listening to you. And this uh paper is the first time uh I am hearing about voice biomarkers digital voice analyzing digital voice for early detection of Alzheimer's. I love it. like biomarkers in general. I'm I come more from the tissue biomarker world, but in general biomarkers of disease now with um AI probably like if you're typing, although now everything is going to voice, it


00:21:46
could probably monitor your typing and if you're declining, it could do that too. Um it's just me imagining all the possibilities. But let's let's dive into this paper. Um what they did here this is a group from uh US mass general California different universities in California uh basically a long long list of authors from the US and um here early detection of Alzheimer disease is critical yet challenging particularly in younger individuals And I know younger well younger like not


00:22:32
even yet retirement age individuals that are suffering from Alzheimer's. And I recently um heard a story of a person that was 37 when they when they started being sick. So, uh, that's crazy to me because in my mind it was like old people's disease. H, and no, not necessarily. So, uh, this study leverages AI to analyze voice recordings from craft story recall task. So this thing craft story recall task is a specific um like neurocognitive task where you hear a story then you um are recalling it


00:23:16
immediately and then the next step is to recall it after 20 minutes and they are checking okay how much did you remember how much did you remember like verbatim literally with the same words from the story how much did you rephrase and and that's like an official neurocognitive test. So they did that um and there was this longitudinal early onset Alzheimer disease study and they used it to detect cognitive impairment and differentiate a early onset AD from early onset nonAD cognitive impairment. Um and they used


00:23:53
speech samples from 120 patients and 68 cognitively unimperred controls. uh and they employed two classification approaches. First feature engineineered machine learning. So they already like knew what they were looking for and they could uh engineer these features and then they had this end to end deep learning incorporating large language model um and they measured the performance and to detect mild cognitive impairment. the feature engineered model um it was using acostic and linguistic features achieved AU of 0.945


00:24:34
um in the test set and then the end to end deep learning model was slightly higher 0.98 but both super high and then for differentiation of early onset Alzheimer from early onset non-Alzheimer uh the featured engineered was engineered model was AUC 0.8 and the end to end model was 0.9. So, uh it was a bit better and explanability analysis revealed reduced lingui linguistic informativeness as a key Alzheimer disease indicator. And I checked for you what this is. Um you basically simplify the way you speak. Um and


00:25:21
it's like a like a feature. Sometimes you use it on uh purpose uh to communicate better, right? But uh if this is like a pattern in your speech in as this specific um craft story recall task, then uh this was an uh indicator of Alzheimer disease. I thought this was fantastic. Um, and I'll have a lot of voice data. So, I don't know if there ever is something like the um, if I I don't know, can you even on your own start getting worried about you getting Alzheimer's or is it people in your


00:26:00
surrounding that notice something that you're like start repeating questions and things like that? Anyway, I do have uh, this device. If anybody is interested in this, send me an email or send me a LinkedIn message. I'm going to send you the link to this. Uh it supercharges your um like email writing, brainstorming, uh insights and things like that. It's called Plaude Pin. And let me check a few comments here. Um I disappeared. Sorry guys. Let me reappear myself. Okay, so we have this question and why I


00:26:43
don't know why the why the font is so huge, guys. Um, give me one second. Let's just do this one. So um application specialist in the field of cytogenetics IHC fish digital pathology for years we often hear the demand uh on of traditional fish um will slowly go down with emerging optic genome mapping honestly I cannot answer this question because I don't know the answer but um like the trend I see in this is maybe there is going to be an easier method method and I'm I I don't know how


00:27:34
optical genome mapping works but uh basically what I see here is okay there is another method that can replace an older method probably in some uh in some instances but what I'm seeing also is that there's a lot of things happening in parallel and a lot of combinations of classical and new cutting edge so um either there's going to be a transition period or there is going to be um things happening in parallel for longer. Um and thank you so much Guido for joining from Argentina and we have a question about


00:28:21
about my device. I'm going to uh I'm going to send a link to this to my list. So again, who is not on my list? I'm going to give you another chance to join the list through the book. Um, and we have one more paper. Let me check. Uh, I'm just going to send you what it is, how I use it, and I um I'll send you the um affiliate link that I have for this, but or I should make a video. I should make a video. Let me know in the comments if I'm supposed to make a video about that.


00:28:52
How I use it to supercharge my productivity. I'm just exploring. So I supercharged my productivity maybe like four times, seven times. That's already enough times, right? Okay, let's do next paper. But I need to remove my code. Okay. Prognostic prediction in soft tissue sircomomas using deep learning and digital pathology of tumor and margin areas and this one is pretty interesting and I'm going to tell you why. Uh this is a group from France, different institutions in France and um here they


00:29:43
say they hisystological FNCC. This is a French um French diagnostic body that is um defining what these uh criteria are for for uh prognostic factors in soft tissue sarcomomas. And it um relies on the differentiation of the tumor, mitoic figures and necrosis. H and this is supposed to be be a prognostic factor, right? Um but it fails to fully capture high-risk patients. And this study aimed to develop a validate a deep learning model to predict metastatic relapse-free survival using digital H& stained horse light images.


00:30:27
So here we are predicting um predicting uh MFS uh metastatic relapse free survival based on H& um as it was a retrospective analysis and was conducted in 308 soft tissue saroma patients in two cancer centers and there was a training cohort of 19 49 patients and two independent validation cohorts of uh 64 and 95 patients. And they used so-called multi-instance learning convolutional neural net um to um and they it they trained it on distinct tumor regions to distinct tumor region centers periphery


00:31:13
and margins to optimize predictive performance. And they found out that models using t tumor center and periphery or their combination were consistently associated with metastatic relops free survival but not really uh the models incorporating margins. Uh so they demonstrated less uh reliable associations. Why why do I think this is important uh or interesting? Because here we're talking about uh a different type of uh tumor soft tissue saroma it's a messeno tumor and in epithelial tumors


00:31:53
for example colurectal cancer the tumor margin or tumor invasive front is like the battlefield uh for the things happening on the immune oncology front right you want to quantify things in the tumor margin and here in this saroma soft tissue saroma study they They showed that the margin is not that important but periphery and center uh can help you uh predict things. And um they confirmed that high-risk score from um deep learning models were independent predictors of um of the MFS and the deep learning CP which is the


00:32:41
center and periphery outperformed their usual grading in prognostic accuracy. um and tumor margin information did not improve predictions. So that was interesting. I thought that you know you have different biology different origin of the tumor and different regions of the tumor are important and so now in the u cloud of citity and immunofllororescent and biomarker like very precise biomarker analysis I would like to know okay what what is expressed in these in the margin versus the periphery versus


00:33:25
the center uh just like to learn about the biology. It's interesting because here they were uh working on this from H& right they can predict metastatic relapse free survival and now the question is why how you can go to like higher plexing methods and check it because you already know where to check uh and that there is something happening. So I love like the there there is the synergy of methods not one of the other but one insight from one from AI on H& leading into a potential to using something else.


00:34:06
Um other than that next week we will not have our live stream because it's in the US it's the Thanksgiving time. So, people are going to be giving thanks and I will give you thanks as well. First, uh offering you the book again. So, if you're not on my list, you're going to get on my list when you get the book and then you're going to be informed about everything, all the podcasts, all the uh I don't know, promotions, offers, what's new in the store. Uh if you want to


00:34:42
thank somebody with beautiful earrings or buy them for Christmas. Why is it not focusing on my beautiful earrings? Come on. Then uh you do want to go to the store. I'm going to show you the code for the store as well. It's going to be in the in the corner. For me, it's the upper left. For you, I don't know. So, um, today I have cartilage, but um, instead of my favorite multi-ucleiated giant cell. Come on. Hey, screwed up my camera. Come on. Okay, now this is a multi-ucleated giant cell. My favorite type of cell.


00:35:22
Uh, it's a bunch of macrofasages together fusing. And then we also have the we already showed these and I have cartilage. We have these um and if you're on the list, there's going to be a little thank you something coming uh in the email. So, um be on the lookout for that new podcast coming out next week. And if you're not subscribed and you are watching this on YouTube, it would be fantastic if you could subscribe to the channel. uh that helps us grow that helps us bring this digital


00:35:59
pathology and medical AI message to more people more researchers and uh your support is so important and it's making this whole work work. So my trailblazers you are amazing. Thank you so much for joining me today and I talk to you in the next