Skip to content
DraconiticResearch
Search Sources API
197,688 studies31 sourcesUpdated 22 min ago
Peer-reviewedCC BYNPJ precision oncology

Skill-guided agentic reasoning enables longitudinal synthesis of myeloma records at expert-level concordance.

  • Johannes Moll1,2,3
  • Jannik Lübberstedt4
  • Christoph Nuernbergk5
  • Jacob Stroh5
  • Luisa Mertens5
  • Anna Purcarea5
  • Christopher Zirn4
  • Zeineb Benchaaben4
  • Fabian Drexel6,4
  • Hartmut Häntze4,7
  • Anirudh Narayanan4,7
  • Friedrich Puttkammer4,7
12 more authors
  • Andrei Zhukov8
  • Jacqueline Lammert9,10
  • Sebastian Ziegelmayer4
  • Markus Graf4
  • Marion Högner5
  • Marcus Makowski4
  • Florian Bassermann5,11,12,13
  • Lisa C Adams4
  • Jiazhen Pan6,14,15
  • Daniel Rueckert6,15,16
  • Krischan Braitsch5
  • Keno K Bressem4,17,18,19
19 affiliations
  1. 1Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital, Munich, Germany. [email protected].
  2. 2Technical University of Munich, School of Medicine and Health, TUM University Hospital, Klinikum rechts der Isar, Department of Diagnostic and Interventional Radiology, Munich, Germany. [email protected].
  3. 3Technical University of Munich, School of Medicine and Health, TUM University Hospital, German Heart Center, Department of Cardiovascular Radiology and Nuclear Medicine, Munich, Germany. [email protected].
  4. 4Technical University of Munich, School of Medicine and Health, TUM University Hospital, Klinikum rechts der Isar, Department of Diagnostic and Interventional Radiology, Munich, Germany.
  5. 5Technical University of Munich, School of Medicine and Health, TUM University Hospital, Klinikum rechts der Isar, Department of Medicine III, Munich, Germany.
  6. 6Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital, Munich, Germany.
  7. 7Department of Radiology, Charité - Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt Universität zu Berlin, Berlin, Germany.
  8. 8Department of Gastroenterology, Infectious Diseases and Rheumatology, Charité - Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt Universität zu Berlin, Berlin, Germany.
  9. 9Chair of Medical Informatics, Institute of Artificial Intelligence in Medicine and Healthcare, TUM School of Medicine and Health, Technical University of Munich, Munich, Germany.
  10. 10Clinical Department of Gynecology, TUM University Hospital, TUM School of Medicine and Health, Munich, Germany.
  11. 11TranslaTUM, Center for Translational Cancer Research, Technical University of Munich, Munich, Germany.
  12. 12Deutsches Konsortium für Translationale Krebsforschung, Heidelberg, Germany.
  13. 13Bavarian Cancer Research Center, Munich, Germany.
  14. 14Department of Engineering Science, University of Oxford, Oxford, UK.
  15. 15Munich Center for Machine Learning (MCML), Munich, Germany.
  16. 16Department of Computing, Imperial College London, London, UK.
  17. 17National Center for Tumor Diseases West, Essen, Germany.
  18. 18Institute of Artificial Intelligence in Medicine, University Hospital Essen, Essen, Germany.
  19. 19Department of Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, Essen, Germany.
Read full textPDFDOICite

Abstract

Patients with multiple myeloma accumulate hundreds of clinical documents over years of sequential therapy, and every treatment decision depends on synthesis across that history. Whether large language model-based systems can match expert agreement on this task, and where they fail, has not been established. On longitudinal records of 811 patients treated over 25 years (44,962 documents and 1,334,677 laboratory values), an agentic reasoning system guided by an indexed clinical skill library was compared against single-pass retrieval-augmented generation (RAG), iterative RAG, iterative RAG augmented with the same skill library, and full-context input on 469 expert-adjudicated patient-question pairs, with an exploratory cross-dataset evaluation on MIMIC-IV. The skill-guided agent exceeded the ceiling at which iterative RAG (75.7%) and full-context input (75.8%) converged, reaching 80.1% concordance, with its advantage widening on the most complex tasks (+11.2 percentage points) and longest records. Providing the same skill library to iterative RAG matched the agent in aggregate (79.0%) and on shorter records, whereas the agent retained a significant advantage on the longest records (+10.3 percentage points in the top decile, p = 0.017) and a numerical, non-significant advantage on the most complex questions (+5.4 percentage points). The aggregate error rate (12.2%) was comparable to expert disagreement (13.6%), but a higher proportion of system errors were clinically significant (57.8 versus 18.8%).

Related

  1. Multidimensional Clinical Deterioration in Interstitial Lung Disease: Longitudinal Concordance of Frailty, Patient-Reported Outcomes, and Lung Function in a Multinational European Registry Cohort Study
    Silke Tello, Anita C. Windhorst, Heva Mustafa +7
    RespirationPeer-reviewedCC BY-NC 4.0Full textGeneral medicinePMC
  2. Tumor-Informed ddPCR for Personalized Longitudinal ctDNA Monitoring in Advanced Solid Tumors: Molecular–Radiological Concordance and Molecular Lead Time
    Dadasaheb Akolkar, Sneha Puranik, Sachin Apurwa +13
    International Journal of Molecular SciencesPeer-reviewedCC BY 4.0OncologyCrossref
  3. HealthFound: a health world model for quantitative reasoning on longitudinal health profiles
    H. Guo, L. Song, P. Ren +8
    medRxivPreprintCC BY-NC-NDFull textEpidemiology & public healthmedRxiv
  4. Severity-stratified, longitudinal cis-eQTL mapping enables high-sensitivity detection of candidate causal genes for critical COVID-19
    E. Chen, G. Gabernet, D. B. Corry +38
    medRxivPreprintCC BY-NC-NDFull textHuman geneticsmedRxiv
  5. Assessing the reproductive lifespan of female cancer survivors using AMH: a meta-synthesis of longitudinal studies
    Neerujah Balachandren, Stavroula L Kastora, Richard A Anderson +5
    Human Reproduction UpdatePeer-reviewedCC BY 4.0Full textObstetrics, gynaecology & paediatricsPMC

Cite

BibTeXRIS
BibTeX
@article{moll2026skillguided,
  title = {{Skill-guided agentic reasoning enables longitudinal synthesis of myeloma records at expert-level concordance.}},
  author = {Johannes Moll and Jannik Lübberstedt and Christoph Nuernbergk and Jacob Stroh and Luisa Mertens and Anna Purcarea and Christopher Zirn and Zeineb Benchaaben and Fabian Drexel and Hartmut Häntze and Anirudh Narayanan and Friedrich Puttkammer and Andrei Zhukov and Jacqueline Lammert and Sebastian Ziegelmayer and Markus Graf and Marion Högner and Marcus Makowski and Florian Bassermann and Lisa C Adams and Jiazhen Pan and Daniel Rueckert and Krischan Braitsch and Keno K Bressem},
  journal = {NPJ precision oncology},
  year = {2026},
  month = oct,
  doi = {10.1038/s41698-026-01733-2},
  url = {https://europepmc.org/article/MED/42875467},
  note = {CC BY},
}
RIS
TY  - JOUR
TI  - Skill-guided agentic reasoning enables longitudinal synthesis of myeloma records at expert-level concordance.
AU  - Johannes Moll
AU  - Jannik Lübberstedt
AU  - Christoph Nuernbergk
AU  - Jacob Stroh
AU  - Luisa Mertens
AU  - Anna Purcarea
AU  - Christopher Zirn
AU  - Zeineb Benchaaben
AU  - Fabian Drexel
AU  - Hartmut Häntze
AU  - Anirudh Narayanan
AU  - Friedrich Puttkammer
AU  - Andrei Zhukov
AU  - Jacqueline Lammert
AU  - Sebastian Ziegelmayer
AU  - Markus Graf
AU  - Marion Högner
AU  - Marcus Makowski
AU  - Florian Bassermann
AU  - Lisa C Adams
AU  - Jiazhen Pan
AU  - Daniel Rueckert
AU  - Krischan Braitsch
AU  - Keno K Bressem
PY  - 2026
DA  - 2026/10/14
JO  - NPJ precision oncology
DO  - 10.1038/s41698-026-01733-2
UR  - https://europepmc.org/article/MED/42875467
AB  - Patients with multiple myeloma accumulate hundreds of clinical documents over years of sequential therapy, and every treatment decision depends on synthesis across that history. Whether large language model-based systems can match expert agreement on this task, and where they fail, has not been established. On longitudinal records of 811 patients treated over 25 years (44,962 documents and 1,334,677 laboratory values), an agentic reasoning system guided by an indexed clinical skill library was compared against single-pass retrieval-augmented generation (RAG), iterative RAG, iterative RAG augmented with the same skill library, and full-context input on 469 expert-adjudicated patient-question pairs, with an exploratory cross-dataset evaluation on MIMIC-IV. The skill-guided agent exceeded the ceiling at which iterative RAG (75.7%) and full-context input (75.8%) converged, reaching 80.1% concordance, with its advantage widening on the most complex tasks (+11.2 percentage points) and longest records. Providing the same skill library to iterative RAG matched the agent in aggregate (79.0%) and on shorter records, whereas the agent retained a significant advantage on the longest records (+10.3 percentage points in the top decile, p = 0.017) and a numerical, non-significant advantage on the most complex questions (+5.4 percentage points). The aggregate error rate (12.2%) was comparable to expert disagreement (13.6%), but a higher proportion of system errors were clinically significant (57.8 versus 18.8%).
SN  - 2397-768X
ER  -