Florian Odi Stummer

Dr. Florian Odi Stummer

orcid

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
2 April 2026 in Cambridge Open Engage
Verified
Preprint
Open access
FSFlorian Odi Stummer
Florian Odi Stummer
Self-evolving AI agents in clinical documentation may manufacture distorted disease prevalence at population scale. This paper formalizes the risk as a partially observable decision process (POMDP) where AI observes only administrative state (coded data) while clinical state remains hidden, creating a structural reward asymmetry: administrative rewards are immediately measurable while clinical outcomes are delayed and noisy. We introduce the artificial epidemiology divergence D(C) = P_a(C) - P_c(C), measuring the gap between administrative and clinical prevalence, and propose five experimentally testable markers for detecting population-level distortion. A computational pre-study on synthetic electronic health records (Synthea v3.x, n = 11,475 patients, 415,464 SNOMED-CT coded conditions) operationalizes D(C) for five sentinel conditions. Results show substantial baseline divergence: D(C) = +0.378 for diabetes (4,334 patients coded without supporting laboratory evidence), D(C) = -0.024 for hypertension (1,741 patients with clinical evidence but no code), and D(C) = +0.101 for obesity. Comorbidity co-occurrence ratios exceed expected values by 1.5x to 2.4x across all sentinel condition pairs. A documentation-action gap of 75.7% is observed for coded diabetes (diagnosed but pharmacologically untreated). A complementary governance simulation demonstrates that the reification feedback loop amplifies coding distortion by 17x to 21x over five iterative cycles under business-first and equilibrium governance scenarios. All data are synthetic; these results establish measurement methodology and construct operationalization, not clinical evidence. Validation on real-world EHR data with linked clinical registries is the necessary next step.

29 July 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
FSFlorian Odi Stummer
Florian Odi Stummer
Minimal Medical Data Sets (MMDS), the standardized, coded datasets underpinning healthcare informatics, AI training, and population health analytics, are widely treated as neutral representations of clinical reality. This paper argues that MMDS entries are not recorded but constructed through a continuous negotiation among three competing forces: clinical diagnostic pressure, billing and audit pressure, and documentation burden. We present a theoretical framework, not an empirical study, that characterises this clinical-administrative negotiation and its consequences for downstream AI and policy systems. Drawing on the sociology of classification (Bowker and Star, 1999), medical science and technology studies (Mol, 2002; Timmermans and Berg, 2003), and documentary theory (Daston and Galison, 2007), we develop the concept of documentary enactment and demonstrate its structural manifestation in the German Federal Health Data Center (FDZ Gesundheit), where a single patient encounter is fragmented across 28 data tables in five ontological categories serving approximately 75 million insured persons. We identify a reification feedback loop in which documentation requirements generate coded data, coded data inform policy and AI training, and policy creates new documentation requirements, progressively embedding administrative categories into the clinical record. We introduce the concept of artificial epidemiology to describe the systematically biased population health picture that emerges when AI systems train on these negotiated artefacts: spurious comorbidity associations from co-occurring billing codes, under-representation of hard-to-code conditions, and over-representation of high-reimbursement diagnoses. These claims are subject to important limitations: this is a single-author theoretical analysis without primary empirical data, situated primarily in the German and European healthcare context, and the illustrative clinical vignette is a composite construct, not derived from patient records. The framework has implications for AI training data governance, health data interoperability policy, and the emerging European Health Data Space (EHDS), which enables cross-border secondary use of MMDS-type data at continental scale without standardised mechanisms to indicate the documentary context in which coded entries were produced.

1 June 2026 in Cambridge Open Engage
Verified
Preprint
Open access
FSFlorian Odi Stummer
Florian Odi Stummer
Background. Clinical AI systems routinely score notes against curator-authored canonical-symptom lists. The dominant rule, k=1 set-membership matching, is rarely evaluated under encounter-complexity stratification. We pre-registered a substrate-level test of the Inverse Care Law direction at the data layer: high-complexity encounters were predicted to show higher misalignment than low-complexity ones. Methods. We computed a per-encounter complexity-density score across 697,021 synthetic encounters (Synthea seed_42_n10000) and stratified adjunctness outcomes by complexity decile against a pre-registered falsification protocol. Adjunctness was scored under two Mapping-Curator lookups (hand-curated SNOMED, n=73; MEDLINEPLUS-derived, n=66) and four locked weight regimes including a zero-history sensitivity probe. The pre-registration locked the score, decile boundaries, four falsification triggers, and nine predicted bands. Results. The predicted direction was falsified. SNOMED MISALIGNED rate fell from 63.56 % at decile 1 to 19.98 % at decile 10 (Spearman rho = -1.000; Delta = -43.58 pp). MEDLINEPLUS fell from 81.19 % to 58.73 % (rho = -0.903; Delta = -22.46 pp). Three of four triggers fired in the inverse direction. The mechanism is condition-side LOOSE-rescue: encounters with more documented conditions accumulate more rescue paths, so notes that would otherwise score MISALIGNED are reclassified as LOOSE. The gradient is robust across all four weight regimes. Conclusions. Standard set-membership symptom-adjunctness scoring rewards documented complexity by mechanically expanding the rescue pool, producing inverted Tudor-Hart algorithmic-forgiveness inequity at the data layer before any human decision. The finding generalises to any k=1 construct whose canonical-symptom-set size grows with comorbidity count. We recommend complexity-stratified evaluation with pre-registered directional hypotheses before deployment.

21 May 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
FSFlorian Odi Stummer
Florian Odi Stummer
SFSebastian Fuller
Sebastian Fuller
The Loss Distribution Approach (LDA), which underpins operational-risk capital quantification under Basel II/III and Solvency II, assumes that loss events are independent and identically distributed. Multi-tool administrative artificial intelligence deployments in healthcare violate this assumption in a topology-specific manner. Drawing on the Cascade Amplification Factor (CAF) framework and empirical simulation data from a recently published five-node primary-care AI pipeline (Stummer 2026, Cambridge Open Engage preprint), we show that three topologies of the same five-tool deployment produce materially different risk regimes: two adjacent AI nodes exhibit subadditive accumulation (CAF = 0.503, 95% confidence interval 0.458 to 0.540), five AI nodes in linear configuration produce near-additive accumulation (CAF = 1.009, 95% CI 0.925 to 1.099) compatible with the LDA assumption, but the same five nodes with a single referral-to-scheduling feedback loop produce superlinear accumulation (CAF = 2.245, 95% CI 2.032 to 2.486), exceeding the pre-specified clinical-relevance threshold by a factor of 1.87 and violating the LDA independence assumption. The transition from the compliant to the non-compliant regime is driven not by the number of AI tools but by the addition of a single feedback edge. We propose a topology-sensitive adjustment of the standard LDA formulation, Capital_adjusted = Capital_LDA × max(1, CAF), analyse the capital-requirement consequences for three illustrative healthcare-provider configurations, and discuss three non-exclusive regulatory pathways: soft-law guidance from the European Insurance and Occupational Pensions Authority, targeted amendment of the Solvency II delegated acts, and market-led standardisation through pilot pricing by specialty operational-risk carriers. Limitations: the CAF values are derived from a single published simulator; the proposed adjustment is a varianceinflation proxy rather than a full copula-based dependency model; capital-requirement estimates are order-of-magnitude; and the regulatory analysis is focused on the European Union, with United States and Asia-Pacific jurisdictions analysed only in outline.

14 April 2026 in Cambridge Open Engage
Verified
Preprint
Open access
FSFlorian Odi Stummer
Florian Odi Stummer
Healthcare cybersecurity concentrates on two threat classes: data confidentiality breaches and adversarial perturbation of machine learning models. This leaves a third attack surface unexamined: the ontological layer, the classification systems (ICD-10, CPT, SNOMED CT) through which clinical encounters become structured data. This analysis is limited to administrative AI in primary care contexts. Recent deployment of administrative AI has demonstrated that this layer is already subject to measurable distortion: documentation tools inflate symptom levels across all six RDoC domains (estimated increases of 30--51%; Castro et al. 2026) while reducing clinical interventions (adjusted OR 0.83), and coding assistants shift evaluation-and-management levels upward by 8--13 percentage points. Each documented distortion mechanism constitutes a potential attack vector. This paper presents a six-class threat taxonomy for ontological attacks on clinical AI systems: (1) Ontology Poisoning, (2) Cascade Injection, (3) Semantic Confusion Attacks, (4) Documentation Flooding, (5) Knowledge Supply Chain Compromise, and (6) Feedback Loop Exploitation. Each class is characterised by attack surface, access level, detectability, harm profile, and analogous traditional attack. I argue that existing security frameworks (NIST CSF 2.0, MITRE ATT&CK) and regulatory instruments (EU AI Act, NIS2) lack coverage for ontological attacks. I propose ontological integrity, the fidelity of classification systems under AI mediation, as a security property requiring dedicated monitoring. Three limitations bound the analysis: the taxonomy is anticipatory, the threat classes derive from one pipeline architecture, and economic incentives relative to traditional cybercrime remain unquantified.

14 April 2026 in Cambridge Open Engage
Verified
Preprint
Open access
FSFlorian Odi Stummer
Florian Odi Stummer
Background. Primary care deploys small language models (SLMs) in sequence. Whether input representation modulates pipeline fidelity, and whether FHIR wrapping or archetype-based structure is required, is not established at small-model scale. Methods. We establish a three-arm paired simulation protocol. Each of 50 Synthea patients was processed through a Triage → Coder SLM pipeline (Gemma 3 1B, Qwen3 1.7B) under three conditions: flat tabular text, compact FHIR R4 Bundle JSON, and openEHR Composition JSON with four Clinical Knowledge Manager archetypes. Three repeats per condition at temperature 0.5 (450 runs). Compact FHIR was pre-specified to control an input-length confound (full FHIR Bundles run approximately 6.5× openEHR length). Primary outcome: semantic F1 against Synthea condition descriptions. Statistics: Friedman test, pairwise Wilcoxon with Bonferroni correction, paired Cohen's d. Results. Mean semantic F1: Flat 0.193, FHIR 0.198, openEHR 0.228. Friedman significant (χ² = 7.56, p = 0.023). Bonferroni-corrected pairwise: Flat vs FHIR p = 1.00 (d = +0.03); Flat vs openEHR p = 0.10 (d = +0.29); FHIR vs openEHR p = 0.09 (d = +0.33). Best arm per patient: Flat 10, FHIR 15, openEHR 25. Directionally consistent with an openEHR lift; no pairwise contrast survives Bonferroni at n = 50. Conclusions. The protocol establishes a reproducible instrument for representation-format effects in multi-SLM clinical pipelines. The n = 50 illustration gives preliminary evidence consistent with long-standing arguments that ontological richness, not mere structural wrapping, drives semantic preservation through a language model cascade. Findings invite collaborative follow-up on larger, non-synthetic cohorts.