In the latest episode of the Oncology Data Advisor podcast, Kristin Maloney shines a spotlight on Mendel AI’s groundbreaking work in clinical intelligence. Mendel AI is a company built at the intersection of technology and medicine, delivering purpose-built solutions designed to empower clinicians, researchers, and decision-makers in healthcare.
Kristin emphasizes the transformative potential of Mendel’s innovative tools, such as Retina (OCR), Resolve (De-identification), and Hypercube (clinical data analytics), in tackling some of the most pressing challenges in modern medicine. Mendel’s proprietary technology is designed not only to process massive volumes of structured and unstructured data but also to understand it at a level that mirrors the reasoning of a seasoned clinician.
One of the key takeaways from the discussion is Mendel’s ability to:
Kristin also underscores Mendel’s dedication to democratizing healthcare AI by making their solutions approachable for both technical and non-technical users. By combining symbolic AI and machine learning, Mendel ensures that its clinical intelligence systems deliver consistent, explainable, and actionable insights, addressing key gaps in the current healthcare landscape.
The podcast serves as a testament to Mendel AI’s vision: enabling clinicians and healthcare organizations to ask and answer critical questions about patient care faster and more accurately than ever before. With a steadfast focus on compliance, usability, and clinical relevance, Mendel AI is poised to redefine how healthcare professionals interact with and derive value from their data.
[00:00:00] Waqas Haque: Hi, everybody. Welcome to our episode of the Oncology Advisor podcast, the latest AI technology episode. Today, we're joined by Kristin Maloney as part of Mendel, which is an AI platform designed to mimic physician level cognitive abilities. Kristin is a certified oncology nursing clinical informaticist with over 15 years of experience, in bedside nursing, clinical research, and informatics.
[00:00:19] Thanks so much for joining today, Kristin.
[00:00:21] Kristin Maloney: thanks for having me. I'm excited to be here.
[00:00:23] Waqas Haque: Yeah. tell us a little bit about, what inspired Mendel and the company's vision.
[00:00:27] Kristin Maloney: Mendel was co founded in 2017 by an AI scientist and a physician who shared this vision of creating a Jarvis for medicine, they wanted to go beyond just being another AI company, making grand claims without substance.
[00:00:42] to achieve this, Mendel spent several years focused on rigorous R& D before launching our technology in 2020. And today, Mendel is one of the most funded clinically I startups with a team of over 100 professionals blue chip investors several industry leading clients and a core [00:01:00] philosophy that we've taken a Mendel is that clinically I isn't just any AI.
[00:01:05] Generative AI, particularly large language models, have transformed many fields rapidly in the recent couple years. there are specific challenges in healthcare. LLMs often struggle with longitudinal reasoning. And lack the deep domain understanding needed because they don't inherently know medicine.
[00:01:25] they entirely lack built in logical reasoning and grounding mechanisms, which are critical for using AI on clinical data. Mendel addresses these gaps with our HyperCube platform, which combines LLMs with a proprietary clinical hypergraph, creating what we call a Neurosymbolic AI system.
[00:01:44] And by anchoring our eye in that medical knowledge, through these knowledge graphs, we can process complex multi-source clinical data at scale, allowing us to unlock actionable insights from clinical data in a novel way. So essentially [00:02:00] that's what the company has set out to do.
[00:02:01] Waqas Haque: Thanks for sharing that.
[00:02:02] And you know, as you mentioned, I mean healthcare, non healthcare, obviously very different settings for applying ai. And I think as, as a clinician myself and you, with your, experience clinically, I think, I think one of the biggest things we challenge. struggle with, you know, clinically in terms of using AI as unstructured data.
[00:02:15] So, you know, you as, as a clinical informatics lead for Mendel, how do you think about this and how does Mendel try to solve this?
[00:02:21] Kristin Maloney: Yeah. So it's true. The unstructured EMR data, such as the physician notes and pathology reports, pose significant challenges for traditional NLP tools.
[00:02:32] those approaches often miss nuanced context due to, the lack of standardization across Documents, authors, and different EMR systems, as well as other data sources. the gold standard approach has been manual human based abstraction which, is slow and can be error prone, is resource intensive, and, you know, there's a huge lag time there.
[00:02:55] It doesn't allow us to get closer to real time insights from the unstructured [00:03:00] data. So our system addresses these gaps in several ways. we consolidate the structured and unstructured data into a single comprehensive record for analysis. we use optical character recognition, OCR, and de identification technology to pre process the documents.
[00:03:18] we also use named entity recognition, NLP, and machine learning to interpret the text. the real differentiator, what I've started to allude to is this clinical reasoning engine, which holistically analyzes The data across the entire record, not just at the document level, this approach enables us to have a deeper, more connected understanding of that longitudinal journey, unlike many other NLP technologies that are usually limited to a document level of reasoning.
[00:03:47] Waqas Haque: Got it. And just to build upon the stuff you're saying about the improved reasoning, how do you use the hypercube to improve the processing time of processing these unstructured documents in the EMR, while also saving the cost per query compared to traditional [00:04:00] NLP?
[00:04:00] Kristin Maloney: it's a great question because we're talking about volumes of data. traditional systems reindex data at every query runtime, which is slow and resource intensive. Mundell indexes the data once, enabling instant querying through our hypergraph technology. our semantically enabled system uses a query language that we built actually called Eloquent. Which simplifies complex SQL like queries. So, for example, instead of searching ICD 10 or Snowbit CT code ranges, an analyst could type something like lung cancer. Eloquent references our hypergraph to account for all synonyms and associated vocabularies and codes automatically.
[00:04:43] this also drastically reduces the query complexity and the runtime, which is how we lower the cost.
[00:04:49] Waqas Haque: Got it. And then kind of shifting gears towards oncology, one thing that we really struggle with in this field is enrolling patients in a clinical trial. A lot of trials fail to meet the recruitment targets.
[00:04:58] many trials even fail to recruit a single [00:05:00] patient. Some cancers, you wait for a trial until you're sort of third, fourth line and out of options for standard of care. and some cancers that are very, aggressive, maybe think about a clinical trial from the beginning. So, how can you use, Mendel and some of the tools like the hybrid cube cohort to improve accrual and recruitment for trials.
[00:05:16] Kristin Maloney: Yeah, it's a great question, and this one is certainly, close to my heart, having worked as a phase one clinical research nurse before, and really experience the challenges with timing and getting the right patients and getting patients access to the right trial at the right point in their care.
[00:05:32] it's a huge challenge. I think there's still a lot from an industry standard level that we want to do to improve this, right? So hypercube cohorts is one of the features in our platform. It enables a lower no code cohort creation. this can be used for clinical trials. It can be used for registries, retrospective studies, other research projects.
[00:05:55] So there's a wide range of applications of creating cohorts and clinical trial accrual is one [00:06:00] of them. Earlier this year, we published an automatic cohort retrieval study, which demonstrated our system's superior speed accuracy and scalability for cohort retrieval compared to models like BERT and GPT 4 with significantly fewer hallucinations.
[00:06:18] So putting that into the clinical trial recruitment workflow, cohorts could be used, for example, when you might need to rapidly find a patient to prescreen to backfill a slot where, you had a patient screen fail, or you need better tools to enable your research team to do more proactive prescreening to help you find patients, for those slower enrolling studies, this is a fast and accurate way to find patients from the sponsors standpoint as well, you know, the solution can help them in their sites make more accurate data driven decisions about feasibility for their trials.
[00:06:53] So, using that full criteria on that patient population without getting a ton of [00:07:00] false positives, they can make better. projections and plan better for the feasibility of their study understand how many sites do we need to open and realistically, how many patients is each of these sites going to be able to put on this particular study and at what pace.
[00:07:15] Waqas Haque: thanks for sharing all that. what should oncologists know about Mendeleyi's sort of ontology and reasoning system?
[00:07:20] Kristin Maloney: Yeah, this is really fascinating, and coming from the clinical world and into AI, it's a different landscape and a different language.
[00:07:30] So let me unpack it a little more with the reasoning system and the ontology for an oncologist. As I explained earlier Mendel's Hypercube mimics how an oncologist reasons. So our custom built ontology that Mendel's built out is focused on the oncology domain, and it organizes clinical concepts hierarchically, integrating all these vocabularies like ARC, SNORM, CPT, ICD 10, SNOMED, and capturing synonyms and [00:08:00] abbreviations.
[00:08:00] what we've done is, we've broken down the clinical concepts in medicine the way you would think about a periodic table of elements. for example, with non small cell lung cancer, it's a neoplasm residing in the lung of a person's body, so you have a tumor, a body site, and this also carries a predefined group of histologic subtypes, so, yeah, adocarcinoma, squamous cell carcinoma, large, large cell, etc.
[00:08:27] so that allows us to understand these clinical concepts the way a clinician would when we read something like non small cell lung cancer in a record. our system boosts the confidence of clinical facts when Corroborating evidence is found. So if a record shows a wedge or section of the left lower lobe, and we don't know anything else, but then we see something that says there's an ICD 10 code in the record for a malignant neoplasm of the lung, these systems can connect these facts together.
[00:08:59] [00:09:00] And boosts our confidence that both are true and kind of validate the accuracy, which helps us to, solve for conflicts and errors in a record, which are actually not uncommon. so then we do this kind of corroboration across dozens to hundreds of clinical facts in the record before producing the final output.
[00:09:20] the last thing I'll say about the key features of the ontology and the reasoning system is we have this equivalency reasoning. for instance, our AI knows that triple negative breast cancer implies negative results for HER2, ER, PR biomarker tests. That all seems very obvious to a clinician, but an AI doesn't necessarily inherently know those things.
[00:09:43] And those are things that we've built out for AI to understand and use so that interconnected understanding in the AI allows us to generate more coherent and actual insights grounded in that clinical context.
[00:09:58] Waqas Haque: Got it. I definitely agree [00:10:00] ontology is very important for oncology.
[00:10:02] How do you think about clinically responsible AI and making sure it's used ethically, especially with patient care.
[00:10:08] Kristin Maloney: at Mendel, we believe that clinically responsible AI must avoid hallucinations, especially when working with very sensitive clinical data and the impacts mistakes, could have on a patient.
[00:10:21] we achieve that by grounding the AI medical knowledge through the reasoning system and ensuring our outputs are both accurate, but also accessible. So as far as explainability goes, this is another thing that I think Mendel has done a really good job at from the outset. So Mendel's platform has been designed from the beginning to deliver a fully explainable AI.
[00:10:41] So it links exactly where the AI got the clinical facts from in each patient record. this is kind of another aspect of how we think about clinically responsible AI. essentially showing our work and making it transparent and easy to validate by a human end user [00:11:00] before they use it. and I think this is an interesting discussion because we're trying to find ways to evaluate if AI is safe and, ready for use in clinical care settings.
[00:11:11] And I think it's an interesting discussion, because we tend to want to benchmark the performance of an AI tool or system to human performance We ourselves at Mendel have done that. We've tried to evaluate our AI quality against our own, human experts. And what we have to recognize is that humans are imperfect too, and they can make mistakes and experience fatigue for sure.
[00:11:36] And, you know, they, there are things that they can only do at a fraction of the pace of an AI technology. So, on the other hand, you know, the AI performs best when It comes to clinical reasoning when it can read the entire record ingest all the information and make sense of it holistically, it does this very quickly compared to a human trying to read the entire record.
[00:11:57] the goal isn't to. [00:12:00] Replace situations where clinical judgment should be used, but to enhance the accuracy and productivity of clinical experts working with patients or patient data.
[00:12:12] Waqas Haque: Got it. Thanks for sharing that. And then can you tell us about maybe recent partnership or a study that you all have worked on?
[00:12:18] Kristin Maloney: Yeah. So I mentioned the automatic cohort retrieval paper already. that's one of the key studies. before that we published in 2023, we published a study that demonstrated how coupling symbolic reasoning with language modeling improves the AI's understanding of unstructured clinical text.
[00:12:36] our findings showed that. The hybrid approach significantly enhanced the extraction of medical variables. we found more information, compared to the LLMs alone. this reinforced the value of combining these methods when, building clinical AI.
[00:12:53] that was a foundational and critical, research finding for us to confirm the hypothesis that Mittel started [00:13:00] with at our outset.
[00:13:01] Waqas Haque: Got it. And then, final question, you know, we're seeing AIBs for so many things, just for you personally, what are some of the really cool use cases you're seeing in terms of applying AI.
[00:13:08] Kristin Maloney: yeah, there's a lot of exciting things coming up. one of the things that we are going to be doing, on a continuous basis, we have an ongoing scientific collaboration with the, the academic, Partners at the University of Pennsylvania it's been great working with them they conducted an initial study to streamline clinical data abstraction in a cohort of patients with either non small cell lung cancer or colorectal cancer the results were presented at ASCO 2024 in an oral abstract session, and they demonstrated that the AI achieved non inferior accuracy compared to humans while significantly improving speed.
[00:13:45] So there's a manuscript detailing these findings that's, expected to be submitted soon by the team at UPenn. building on that success, they are now planning a follow up study with MUNDL, to explore how AI extracted data can be leveraged [00:14:00] for trial matching. this ongoing collaboration not only allows us to validate our technology, but it also highlights kind of the practical value in advancing, both abstraction workflows, which I've talked a lot about, as well as clinical trial matching.
[00:14:14] And then this past year, it's really shown how hypercube can revolutionize this chart abstraction process. So it's not just about the speed and scale, but it's also about how comprehensive and connected the data become when you consolidate those facts across the record. So, 1 of the features in hypercube that we use to do this is called chart review, and it allows the user to
[00:14:37] Search the comprehensive record. We have a semantically enabled search system, which is incredibly powerful. If you don't know what you're looking for. So you open a patient's record. You don't know anything about that patient. You're trying to understand what their journey is. You might not know what keywords you want to search for.
[00:14:53] So you want to start broad and go more narrow. for example, with a melanoma patient, you might want to search immune [00:15:00] checkpoint inhibitors as a starting point to see like. What did this patient maybe get? Or you might just want to search biomarkers or BRAF to understand what the patient's biomarker status was and then build on understanding their story from there.
[00:15:13] so that's a powerful tool. when you go to that search result, we link the two together. you can seamlessly move between the AI outputs and the patient record to create this workflow that. is incredibly powerful for validating the AI results.
[00:15:30] As well as enhancing data quality, and scaling abstraction efforts in a way that just wasn't possible before. our customers are using this tool to rapidly generate high quality verbal data, conduct retrospective cohort studies with far less manual effort. And we also envision this will be powerful as a tool for streamlining processes like medical billing coding and prior authorization.
[00:15:54] And then also, as I mentioned before, it does have incredible potential for finding and then also [00:16:00] doing some of the prescreening activities to identify patients for trials. what's exciting about this. tooling and supercharging these workflows is this kind of transforms the way that we can think about cancer research.
[00:16:14] So a recent example that stood out to me was a research team in an academic center that's working on this retrospective study. And in the study design, they opted to only abstract the first three doses of chemotherapy for a large cohort of patients.
[00:16:28] limiting that scope was because they want to account for what they can do as a team, they don't have forever to focus on abstraction. They've got to get to data analysis. we see this all the time, right?
[00:16:39] With a study focused on what could be achievable within the confines of a human only chart review workflow. So with hypercube. Mendel's system extracts all the medications, whether they're cancer related or otherwise, automatically. this saves significant time and effort, and this kind of approach isn't just, [00:17:00] powerful, in theory.
[00:17:00] I think it can really change how we think about study design and data collection, and even routine interactions with clinical records when we have this kind of power with the AI.
[00:17:13] Waqas Haque: thanks so much for answering all the questions and for our conversation today. All the best with Mendel and, have the rest of your week.
[00:17:19] Kristin Maloney: Great. Thank you so much for having me.
To learn more about Mendel AI’s transformative work and hear Kristin Maloney's full insights, listen to the podcast on Oncology Data Advisor here.
