Amy Carpenter
November 2026—Artificial intelligence has enormous potential to change hematology practice, but achieving that potential will require more than better algorithms.
What’s needed and how to move the field forward is what Amrom Obstfeld, MD, PhD, chair of the International Council for Standardisation in Haematology, spoke to at the ADLM meeting this summer.
Work on interesting applications in hematology is underway, he said, noting a couple of examples, and the AI possibilities look promising. In Dr. Obstfeld’s view, it appears a bottleneck is behind more machine learning not yet having made its way into hematology laboratories.
He cited a report of an international survey of laboratory professionals experienced in AI, developed by the International Federation of Clinical Chemistry and Laboratory Medicine committee on AI in laboratory medicine (Yang HS, et al. Clin Chem Lab Med. 2026;64[7]:1513–1526). “Interestingly, they found that within this selected group of enriched AI-interested individuals, uptake of AI is pretty small. Only a third of them could identify an application that was operational in their own laboratories,” said Dr. Obstfeld, associate chair for pathology informatics and director of coagulation in the Children’s Hospital of Philadelphia core laboratory, and associate professor of clinical pathology and laboratory medicine, Perelman School of Medicine at the University of Pennsylvania.
“It does seem like there’s been perhaps a slowness to implementation,” he added.
The aim of the survey was to assess the current application status of AI/machine learning in clinical laboratories, identify adoption challenges, and collect expert opinions on key implementation issues.
There were 50 complete responses to the survey: 18 from Europe, 18 from the U.S., seven from Canada, six from Asia, and one from Africa.
The authors report that even among laboratory professionals interested in AI, implementation is limited and applications highly heterogeneous. Most clinical laboratories lack the in-house expertise to develop their own models. More external providers are producing AI systems designed to be integrated into clinical workflows through the electronic health record and laboratory information system, and the “emerging trend is toward adopting external models, much like laboratories now use commercial immunoassays,” not developing them but following established verification protocols, the authors write.
They identified the leading challenges as insufficient information technology infrastructure and a lack of practical implementation guidelines.
Additional evidence of a bottleneck, Dr. Obstfeld said, comes from Food and Drug Administration data. As of June this year, only 21 of the 1,525 FDA-authorized AI-enabled medical devices are in the field of laboratory medicine, which is about 1.4 percent of the total, down from 1.9 percent in 2024, he said. He cited examples of hematology AI: Athelas Home hematology analyzer, Scopio X100/X100HT digital blood smear system, and CellaVision analyzers.
Then there’s the published literature. Dr. Obstfeld used AI to create an application to query the literature for mentions of topics related to AI/ML in coagulation, laboratory hematology, and general laboratory medicine journals. The AI-concept density in the general field of laboratory medicine, he said, is more than double that of laboratory hematology and coagulation.
The report of the survey findings, the FDA data, and the results of the literature search, combined with Dr. Obstfeld’s personal anecdotal experience, point to what appears to be a bottleneck, he said, in the use of AI in laboratory hematology and coagulation.
The questions are why and whether there is a role for the ADLM and the ICSH.
Dr. Obstfeld shared his own “gestalt,” as he put it, for how AI can go from ideation to application. The eight steps in his AI implementation pipeline are as follows: clinical question, data access, model development, transparent reporting, external validation, AI governance review, local implementation, and monitoring.
“There are substantial roadblocks and bottlenecks at each and every one of these steps,” he said, from a lack of AI-experienced staff to launch the process to the laboriousness and difficulty of the data extraction process, all the way to workflow and LIS integration barriers and real-time data export challenges that hamper local implementation. “Our analyzers were built to interface with the LIS, not with AI models,” he said.
The two types of AI are homegrown and vendor-purchased or vendor-supported, and Dr. Obstfeld believes even the latter—implementing a vendor solution—is difficult. It starts with a clinical requirement and is followed by vendor identification, model specification review, data flow and integration review, local validation, AI governance review, local implementation, and monitoring.
“The problem,” he said, “is that there are no agreed-upon standards for how a vendor AI solution should move from selection through validation, integration, implementation, and monitoring.”
How can organizations like the ICSH and the ADLM be of help with the challenges of AI standardization? Dr. Obstfeld uses an analogy from the shipping industry, in which containers have been built to ISO standard dimensions of 8 × 8.6 × 20 feet. Any crane, ship, truck, or rail car worldwide can move and stack any container of these dimensions. The ISO standard, in place since 1968, decoupled the cargo from the transport.
“There’s nothing magical” about the dimensions, he said. “The point is that if we all agree on the rules of the road, then we can get past some of these roadblocks that don’t necessarily need to be there. That is the role of standards in facilitating our ambitions.”
The role of the ICSH, as he sees it, “is to remove some of the ambiguity by setting down the rules of the road for how we’re going to do this.” Clearer guidance would help laboratory directors understand how to validate an AI model for their own clinical use. It would help instrument and AI vendors understand what laboratory directors will need from them before a tool is adopted, and it would help data scientists develop models for clinical use that generalize across sites and populations. “And our regulators, who have not yet filled the gap here, need a bit of a push for them to understand what our expectations are.”
ICSH practice standards, such as those that answer questions where laboratory directors or data scientists would otherwise have to guess, could help. Among them:
- How to validate an AI tool locally before clinical use.
- How often to revalidate after a reagent lot change or model update.
- What raw data and metadata to capture and retain.
- How to stress-test a model (spiked direct oral anticoagulant, emicizumab, heparin, high-titer inhibitor).
- How to risk-tier AI tools by clinical consequence.
- What a vendor must disclose: an algorithm package insert.
- Common terminology and export formats for waveforms.
Having experienced a ride in a Waymo vehicle shortly before the ADLM meeting, Dr. Obstfeld wondered: “What are they doing that we are not?”
“They have distinct levels of [driving] automation that relate to the regulatory category that car falls into—everything from level zero with no automation to five” for the self-driving car, he said.
“If we look at some of our hematology digital morphology devices,” he continued, “we’ll find there isn’t such a clear categorization. One vendor might do some subset of functions; another vendor does a different subset of functions.” Without a shared vocabulary for “how much AI” there is, it becomes impossible to compare multiple digital morphology vendors, he said. Each vendor embeds AI differently—from automated cell detection and image acquisition to WBC AI classification, RBC/PLT AI, differential autoverification, decision support, and marrow/fluids—and there is no common way to declare it.
“So we need better bucketing of the applications we’re using so that when we talk about a digital morphology level three, hypothetically speaking, we all know that means it does the differential but it doesn’t do RBC morphology,” he said.
Also needed are standardized tests. “If we’re looking at two AI applications, how do we know which one is better?” Dr. Obstfeld asks. “There is no standardized test by which we can grade different devices or AI models. So we need open, published data sets that we can share to benchmark against one another so we know when we’re making progress or when we’re just spinning our wheels making another paper.”
One of the problems: “Our data is locked down,” he said. It is locked down at the institutional level and within laboratory instruments, making it difficult to extract data from those instruments. “It’s not a first-tier–level function of these analyzers,” he said.
He defined the following specific problems:
- Raw data are not exported to the LIS: Waveforms; scattergrams; and volume, conductivity, scatter/cell population data are reduced to a single number before the LIS ever sees them.
- Export is laborious: Raw data access is rarely a product feature and requires vendor engagement.
- Proprietary formats: Each analyzer emits its own format.
- Metadata sits everywhere: Reagent lot, calibration, software version, and quality control state live in separate systems and must be rejoined.
Even if the data can be extracted and the AI research done, Dr. Obstfeld said, “it is extremely difficult then to put it into production because you can dump that data out, but getting it out on a routine basis as you’re doing testing is close to impossible.” Nearly all published hematology/coagulation machine learning rests on one-off extractions; almost none survives as a production feed.
The ICSH has a role here, in his view, because this is a standards problem to some degree, though he thinks it’s a bigger problem than that. “And it behooves laboratory medicine organizations to work together to analyze the issues, divide up the problem space, and start solving this issue.”
The ICSH, he said, must partner with others to reduce friction in AI deployment, and Dr. Obstfeld proposes the following division of tasks:
- ADLM: general laboratory medicine practice standards (evaluation, adoption, and monitoring principles).
- Association for Pathology Informatics and partners: data and image interoperability (formats, image structure, the Digital Imaging and Communications in Medicine standard, and the “plumbing”).
- ICSH: hematology/coagulation-specific data requirements (waveform export, scattergram/cell population features, flag metadata, prolonged APTT cause annotation).
Medical AI meets the human workflow in three ways, Dr. Obstfeld said: Automation (it does a task a human already does), augmentation (a human plus AI beats either one alone), and extension (it does what no human can do). Autoverification and autovalidation are forms of automation AI and are common in clinical laboratories.
AI augmentation is “a bit more in the future,” he said, but there is talk of using AI assistants to draft reports, such as a hemoglobin identification report. “This is potentially something that could be of use to our hematologists, hematopathologists, and laboratory hematologists who are doing the signing-out process. It also could help our clinicians who are not within our field and sometimes don’t understand the implications of our esoteric testing.” A data scientist in Dr. Obstfeld’s own laboratory is doing work now to automate the hemoglobin identification process.
Then there are the AI applications (extension) that “go way beyond what we can do ourselves,” he said, citing work related to scattergrams of hematology analyzers, with the aim of finding patterns humans cannot perceive. Others have focused on clot waveform analysis, extracting clinically meaningful information from the clot curves that coagulation analyzers produce.
“The overall trend is that our analyzers contain a large amount of raw data that gets reduced to a single data point result that we provide to clinicians, but that raw data has meaningful clinical implications as well,” Dr. Obstfeld said. “At least that’s the hypothesis.”
The potential applications of machine learning in laboratory hematology are broad, extending from preanalytical/laboratory quality control and CBC-based detection and risk to blood film morphology, anemia and cytopenias, and leukemia and myelodysplastic syndromes. “This is certainly an area of active investigation, although the application is still somewhat wanting, in my opinion,” Dr. Obstfeld said.
He cites two AI examples. The authors of a 2025 article applied deep learning techniques to expand the ability to use clot waveform analysis to make a more definitive diagnosis beyond that of a prolonged APTT (Matsuda M, et al. Sci Rep. 2025;15[1]:32336). They trained a neural network based on multiwavelength waveforms, “and were able to do a pretty good job looking at the raw statistics, at least diagnostically, to identify the cause of a prolonged APTT,” he said.
Dr. Obstfeld noted the study’s limitations: It is a single-center study with no external validation and tied to one analyzer. The authors, some of whom are coagulation analyzer researchers, concluded that such models are “best built in each laboratory.”
In the second example, Eckardt and coauthors in Dresden, Germany, used a deep learning approach to predict the genotype of a malignancy based on the morphology of blasts (Eckardt JN, et al. Leukemia. 2022;36[1]:111–118). “They were able to predict whether a patient would have an NPM1 mutation based on marrow smear images,” Dr. Obstfeld said. “It’s not something that most of us can do visually, but given enough outcome data, this is something that is potentially feasible.”
This, too, was a single-center study with internal threefold cross-validation only and no external validation, Dr. Obstfeld noted. The NPM1 model rests on a limited set of mutated cases, and morphologic clues are hypothesis-generating and not yet independently confirmed.
Digital morphology is where the near future is for AI in hematology, Dr. Obstfeld said, and a first priority is to train more AI subject matter experts in hematology.
“We need to train our young people, or older people, anyone who’s interested, to understand these concepts. And by recruiting and training out more of those folks who are interested, the rest will come,” he said.
Amy Carpenter is CAP TODAY senior editor. Dr. Obstfeld’s comments at the meeting were presented under a partnership of the ICSH and the ADLM hematology and coagulation division.