AUTHORS
Valerio Lupperger, PhD, Leroy Freyler2, Lisa Erdl3, Sven Maschek4, Gregor Hoermann, MD5, Wolfgang Kern, MD, PhD6, Torsten Haferlach, MD, PhD7, Christian Pohlkamp, MD8
AFFILIATIONS
1 MLL Munich Leukemia Laboratory, Munich, Germany
2 MLL Munich Leukemia Laboratory, Munich, Germany
3 MLL Munich Leukemia Laboratory, Munich, Germany
4 MLL Munich Leukemia Laboratory, Munich, Germany
5 MLL Munich Leukemia Laboratory, Munich, Germany
6 MLL Munich Leukemia Laboratory, Munich, Germany
7 MLL Munich Leukemia Laboratory, Munich, Germany
8 MLL Munich Leukemia Laboratory, Munich, Germany
ABSTRACT
Background:
Modern leukemia diagnostics require individualized multimodal workups integrating a broad variety of diagnostic methods. In our high-volume leukemia laboratory, disease-specific standard operating procedures (SOPs) are based on international guidelines and classification systems, while heterogeneous, non-standardized clinical requests pose a challenge. Human experts face increasing cognitive and logistical load when translating complex referral information into coherent, patient-specific and often stepwise diagnostic strategies. Turnaround times and economic pressure are critical, staff shortages are growing. We implemented an AI-based pre-analytical test selection system (TSS) that considers SOPs with guideline knowledge, key facts from the request form, and structured data on the referring institution, to generate case-specific proposals for diagnostic workflows.
Methods:
Two preparatory steps were performed upfront: Off-the-shelf large language models (LLMs) were used to build look-up tables exposing SOPs and clinical-institution preferences in machine-readable form, thereafter reviewed by our lab experts. We formerly reported a multi-model optical character recognition framework that extracts key information from scanned referral documents, such as diagnosis, sample material and requested methods, also followed by human review.
This enables a final, LLM-driven TSS to integrate a stepwise composition of the sample workup proposal: First, the suspected diagnosis is mapped by an LLM to one or more entities of a curated diagnosis hierarchy. Second this mapped diagnosis determines the governing, disease-specific SOP: the combination of diagnosis and disease status identifies a scenario specifying a default action per diagnostic method, refined by sample type (peripheral blood, bone marrow) and blood count. Third, the resulting recommendation is reconciled with the specifications from the clinical request form, and with general processing rules on file for the requesting institution. Finally, this algorithm determines one or more actions per diagnostic method: a specific analysis should be (a) performed, (b) deferred, (c) kept pending for clarification with the referring physician, or (d) not performed.
LLMs used throughout the entire process were Claude Haiku 4.5 and Opus 4.7. After establishing the technical framework, the system was evaluated on a retrospective test set including human-curated workflows of 11,612 leukemia samples.
Results:
The TSS suggests a case-specific analysis setup including cytomorphology (C), immunophenotyping (I), chromosome banding analysis (CBA), FISH (F), and molecular genetics (M). It matched the human-validated workup across all applicable tracks in 9,540 of 11,612 cases (82.2 % exact-case concordance); in even 94.5 % of cases the proposal agreed with the expert controls on all-but-one of up to 7 tracks. Per-method F1 score was 0.986 for I, 0.943 for C, 0.941 for M, 0.907 for CBA, and 0.873 for F. Across all 58,060 method-level decisions, the model over-recommended a method in 1,947 (3.4 %) and omitted an indicated method in 1,149 (2.0 %) cases, totaling 3,096 discordant decisions. 100 out of 142 deviant cases (70.4%), selected at random and reviewed by human experts, turned out to propose an at least non-inferior diagnostic setup. Remaining discordances predominantly reflected rare or complex clinical scenarios. The complete LLM-driven TSS workflow has a median runtime of 5 seconds and thus can be placed ahead of any human review without any delay for the overall process. A fine-tuned version has entered a supervised prospective beta phase in 07/2026, testing between 400 and 500 cases per day.
Conclusions:
AI-assisted pre-analytical workflow orchestration at a specialized, high-volume leukemia laboratory is technically feasible and can meaningfully support individualized, multimodal diagnostic workups. Our system integrates guidelines, disease-specific SOPs and client preferences with information extracted from referral documents. It hereby helps standardize diagnostic strategies and decrease cognitive load on staff, while human review of AI suggestions remains obligatory. This approach may serve as a scalable template for AI-enhanced orchestration of high-complexity hematologic diagnostics and can be further extended to dynamically incorporate new classification systems.
DISCLOSURES
Valerio Lupperger, PhD: MLL Munich Leukemia Laboratory - Current Employment