Qnomx AG (James Creeden)

Comment of Qnomx AG, Basel, Switzerland, on Docket No. FDA-2026-N-7874, "Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback". The full comment is attached as a PDF (12 sections, 11 recommendations); this field summarises it.

Qnomx AG is an early-stage developer of retrieval-grounded software for clinical laboratory workflows that would support a qualified user’s interpretation of somatic sequence variants. We have no device authorised or under review in the United States and take no position on the qualification or classification of any software function. We comment because the Discussion Paper does not develop a genomic example.

Our comment responds to Questions 1, 2, 4, 9 to 14, 16, 19, 21 to 23 and 26, and to Sections IV, V.A, V.B.2 and VI.C and Appendix A elements S.3, E.4, R.2 and A.1. The main points are:

1. Traceability to source evidence should enter the risk framework only as measured, claim-level properties (attribution coverage, attribution fidelity under blinded independent adjudication, critical omission, corpus version disclosure, reconstructability), with no credit where fidelity is unreported, the setting affords no verification, or the user cannot evaluate the sources (Q1, Q2).

2. For retrieval-grounded genomic functions, the usual ground truth is contaminated by construction: where the label is in the corpus the system retrieves, agreement with it measures extraction fidelity, not independent application of the classification framework. We propose three separately reported endpoint families and a blinded-adjudicator comparator where no recognised assertion exists, sources conflict, or the draft extrapolates context (Q10, Q14).

3. Human review is a claim that needs evidence about the configured workflow. We report a sponsor-conducted, pre-specified rater study of a related summarisation function in which inter-rater agreement was low and accuracy was the one dimension on which the primary analysis did not conclude non-inferiority, and we ask that inter-rater agreement be reported for any panel comparator and rubric calibration precede confirmatory studies (V.A, E.4, Q14).

4. A change in the content of an externally maintained evidence corpus, with model and code unchanged, is a change category the Discussion Paper does not name. We propose version pinning, a three-tier impact classification, measured detection-and-propagation intervals as a postmarket parameter, and consideration alongside FDA’s public human genetic variant database recognition programme (Q19, Q21 to Q23).

5. Retrieved-content safeguards should attach to retrieval-grounded devices regardless of the agentic boundary (Q26), and sponsor-conducted benchmarking should remain an acceptable route under pre-specification, independent adjudication and full disclosure (Q10, Q16).

Comment on Docket No. FDA-2026-N-7874

Dockets Management Staff (HFA-305)
U.S. Food and Drug Administration
Submitted electronically via Regulations.gov
Docket No. FDA-2026-N-7874
Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback
Date: 19 October 2026

Qnomx AG is an early-stage software developer in Basel, Switzerland. We are developing software for
clinical laboratory workflows that would support a qualified user’s interpretation of somatic sequence
variants. It is designed to retrieve evidence from curated sources at identified versions; to produce
structured, source-cited draft variant and gene descriptions; and to present draft categorisations under
stated published frameworks (ClinGen/CGC/VICC oncogenicity; AMP/ASCO/CAP clinical significance),
together with therapy-association evidence at the source’s stated evidence level. A qualified molecular
pathologist or clinical scientist reviews and confirms the output and remains responsible for the final
classification and for any clinical communication. It is not designed to recommend therapy. This is a
development programme, not a labelled intended purpose or a regulatory status in any jurisdiction.
Software of this kind performs two distinguishable functions: evidence synthesis and the proposal of
interpretive categories. Each needs its own evaluation, and neither should be evaluated by agreement
with the public database labels the system retrieves (Section 4).

Q1 asks whether traceability of an output to primary sources belongs in a risk framework. Source-linked
output is a design goal of systems like ours, so we would benefit if FDA treated traceability as riskrelevant. We have no device authorised or under review in the United States. We take no position on the
qualification or classification of any software function, our own included, or on any statutory exclusion. We
comment because the Discussion Paper does not develop a genomic example: “genomics” appears once
in the body, in the definition of multimodality (p. 4), and none of the worked examples involve genomic
evidence.

1. General observations
We support the two-axis heuristic and the competency-based structure. Three features of the Discussion
Paper matter most for our domain. S.3 treats presenting “uncertain, outdated, or contested information
with false confidence” as a safety failure and considers deferral “without over-deferring on questions
within its knowledge” (p. 24); in clinical genomics a substantial share of variants have evidence that is
insufficient or in conflict. The two-sided criteria (under- and over-refusal, under- and over-escalation,
deferral and over-deferral) should be retained; one-sided criteria produce systems that fail safe into
uselessness. And the approach under consideration in Section V.A, under which the evaluated object
would be “the final user-facing device, as configured and intended to be deployed” (p. 11), has
consequences for software of our kind that Sections 4 and 5 set out.

2. Q1 and Q2: traceability as a measured property
Q1 lists traceability among four example dimensions and gives it no direction. We recommend including it,
defined so that no sponsor can claim it by architecture.
Citations are cheap. A retrieval-grounded system can attach a real, resolvable source to a claim the
source does not support: a case report over-generalised, germline evidence applied to a somatic
question, a superseded assertion, an unfaithful summary. A dimension defined as “the output has
citations” rewards a property unrelated to correctness. Traceability may also cut the other way: E.4 names
automation bias, where fluency or perceived authority reduces scrutiny (p. 25); a dense citation apparatus
supplies both. Whether traceability lowers or raises the consequence of an incorrect output depends on
whether users verify, which depends on another Q1 dimension, the time pressure of the setting.
We propose that traceability enter the framework as measurements at the level of the generated claim:
• Attribution coverage. Of all clinically substantive generated claims, the proportion that resolve to
a specific entry at a stated version. For a claim of the form “variant X, tumour type Y, therapy Z,
evidence level L”, the source is the entry that supports that combination, not the database in
general.
• Attribution fidelity. Of cited claims, the proportion the cited entry supports as stated, in the
stated tumour context and at the stated evidence level, under blinded adjudication by reviewers
independent of the sponsor. This is the endpoint that matters, and the one a sponsor cannot
produce alone. Fidelity should be read together with coverage, since a system that cites little can
score high on fidelity.
• Critical omission. Of independently adjudicated decision-relevant evidence items, the proportion
absent from the output or present but inconspicuous. For evidence-synthesis functions this is
where directiveness emerges (Q2): a summary that omits contradicting evidence is more directive
than one that presents it, whatever language it uses.
• Corpus version disclosure. Whether the output states the version of each source it drew on.
• Reconstructability. Whether the record preserved for a prior output (the output itself, the model
version, prompts, retrieval configuration and source snapshot) is sufficient to audit how the output
was produced; whether material evidence selection is stable across repeat runs under controlled
conditions; and, where variation remains, whether it changes any material evidence statement,
uncertainty flag or escalation. We suggest this in place of identity of output, which can be met by
suppressing variation.

Traceability should earn nothing where fidelity is not reported under independent adjudication against prespecified acceptance criteria, where the setting gives the user no practical opportunity to verify, or where
the user cannot evaluate the sources.

3. Q4 and the measurement discussion (pp. 7-8)
The Discussion Paper suggests that IVD, measurement and signal-processing functions may sit higher
on the consequences axis because “the user typically cannot independently evaluate the basis for the
output” (pp. 7-8). We agree that this is a real determinant of consequence, and Section IV ties risk
assessment to evidentiary expectations (p. 5).

We ask that this presumption be rebuttable by evidence about the specific device rather than settled by
category. When the output includes a prominently displayed draft category, that category may anchor the
reviewer whatever sits beneath it, and a well-supported evidence set does not by itself show that review
happened. The evidence that should count is the degree of independent review demonstrated in use,
measured as in Section 5. Footnote 13 of the Discussion Paper notes that this consideration seems
consistent with the independent-review criterion of section 520(o)(1)(E).

On Q4, the generalist/specialist distinction is a property of the pairing of output and user. The same draft
is a different risk object for a molecular pathologist who weighs expert-panel assertions against functional
data than for a clinician without that training. There are also benefits of the kind Q3 and the generalist
discussion on p. 8 acknowledge: many patients are treated where there is no molecular tumour board,
and access to specialist-level evidence in those settings belongs in the assessment. The mitigation we
propose is that the evidence a user is least able to evaluate be the evidence the device is most explicit
about: what is contested, superseded, single-source, or not found.

4. Q10 to Q14: benchmark validity in clinical genomics
Q10 asks how performance on a benchmark can be shown to predict real-world behaviour.

The verification problem. The Discussion Paper names contamination of public benchmarking assets
(p. 12, footnote 18; Q10); this is how that problem arises in genomics. The evidence base is public:
expert-panel assertions, population frequencies, curated somatic knowledgebases, and the literature
beneath them. Content derived from these sources is present in the text on which large models are
trained. We do not claim that any model was trained on any release; the training data of third-party
models are generally not disclosed in a form a sponsor or a reviewer can audit, and training cutoffs are
self-reported. Where benchmark labels are drawn from the same public sources a model may have been
trained on, agreement with those labels cannot by itself distinguish reasoning from recall.

Temporal holdout is legitimate and insufficient. Withholding recently adjudicated assertions answers
real questions when corpus, retrieval and protocol are frozen to the same date. It does not on its own
show reasoning. The label is withheld but the evidence that produced it is not: the functional studies and
case series usually precede the assertion. Newly adjudicated variants are enriched for the contested and
the rare, so performance on them is a stress result, not an estimate of performance in use. And
withholding current evidence from the retriever evaluates a configuration that is not deployed; it can
supplement, but not stand in for, evaluation of the configured device that Section V.A would evaluate.

Three endpoint families. We suggest evaluating retrieval-grounded genomic functions on three
separately reported families, against a pinned corpus snapshot.
1) Evidence fidelity at the claim level: coverage, fidelity and critical omission as defined in Section
2.
2) Proposed-category performance: agreement of the draft category with independent adjudication
under a stated framework and version, from the same dated evidence, reported by direction
(over-call, under-call) and severity, and whether the draft flags the evidence as insufficient or
conflicting where the adjudicator found it so (S.3), including where recognised sources disagree
with one another. Agreement with a database label should not be accepted as sufficient evidence
of this. Where the target classification, or the assertion it rests on, is already present in the
retrieved entry, concordance with it measures retrieval and extraction fidelity, not independent
application of the framework. A recognised, version-identified assertion for exactly the stated
combination can serve as the reference for that element, and concordance with it should be
reported as what it is. A blinded adjudicator, who did not see the label and applied the stated
framework to the stated evidence, is needed where no such assertion exists, where recognised
sources conflict, or where the draft extrapolates beyond the assertion’s stated context in tumour
type or evidence level, which is where the interpretive work lies. This is where the verification
problem meets the comparator question in Q14. For this class of function the usual ground truth
is often contaminated by construction; the Discussion Paper names contamination but does not
distinguish this case, where the label sits inside the corpus the system retrieves.
3) The reviewed workflow: whether the qualified user’s confirmation step catches the errors the first
two families reveal, under representative conditions.

Power (Q12). A small sponsor can estimate overall performance, characterise its uncertainty and run
targeted stress tests, and should be required to. What it cannot do alone is establish well-powered
stratified estimates across gene and condition, evidence class and ancestry at once, which R.2 (p. 26)
would call for on ancestry and which the Discussion Paper’s interest in consistent performance across
clinically relevant subgroups (p. 16) would by extension call for on gene, condition and evidence class.
Expert-panel adjudications accrue slowly, and once stratified in this way the per-cell counts can fall to
single digits. We suggest a layered expectation: sponsor evaluation for overall performance and stress
tests; independent adjudication for fidelity and category endpoints; shared sequestered challenge sets,
maintained by third parties (Q16), for the stratified estimates. This adds to sponsor evaluation; it does not
replace it. Where stratified estimates cannot be obtained, the limitation should be stated in labelling.

Synthetic data (Q13). Synthetic genomic inputs from models of the same class risk reproducing the gaps
the evaluation exists to find, and the risk is greatest where real data is thinnest. Synthetic variation is
defensible for format and phrasing (nomenclature, request wording, unit reporting). It should not be used
to establish performance on evidence content unless the reference answers were established
independently of any model of the same class, and sponsors should say which purpose it serves.

Which confirmation approaches apply (Q11). Of the approaches on p. 15, standardised patient
interactions have no object for a laboratory-workflow function. The primary method for the first two
endpoint families is retrospective evaluation on real inputs with blinded, independent adjudication.
Shadow deployment, in which outputs are not shown to the user (p. 15), cannot answer the third family’s
question, which is how reviewers behave when the draft is in front of them. To Q11, prospective
observation of the reviewed workflow, with the draft displayed, may address this. The Discussion Paper
already ties the choice of approach to intended use, risk and the type of data the device acts on (pp. 1415). Future guidance could illustrate that tailoring with a laboratory-workflow example, so that the
“approximate order of increasing rigor” (p. 15) is not read as a ranking of informativeness across classes.

5. Section V.A and E.4: human review is a claim that needs evidence
Software of our kind is designed for a qualified user who reviews and confirms its output. It would be easy
to treat that reviewer as a mitigation and stop. For a device whose safety argument rests on review, the
configured device includes the interface and the working conditions. A traceable evidence chain the
reviewer cannot inspect in the time the workflow allows is a property of the document, not of the system in
use.

Where a sponsor relies on human review, it should show that the configured workflow supports it: time to
inspect the cited basis of a substantive claim under representative conditions; whether contradicting
evidence is prominently presented; whether reviewers detect seeded errors of the kinds the system
actually makes, by seniority and case complexity; whether citation density or polished prose raises
acceptance of incorrect content; and which output categories the system should decline to draft because
review cannot be relied on to catch a failure.

Our recent study bears on comparators (Q14). It was a blinded, counterbalanced rater study of a related
summarisation function, conducted with a hospital research institute as a clinical performance study
documented under ISO 20916:2019 (we assert no classification of the system evaluated), with a design
and analysis plan pre-specified and published before data lock, and reported as a preprint [1]. Three
qualified oncologists scored AI-assisted and manual summaries of 37 tertiary pediatric oncology
genomics reports (74 paired evaluations) against the source reports on a published six-dimension, fivepoint rubric. Both co-primary composites met non-inferiority at the pre-specified 0.5-point margin: content
+0.05 (95% credible interval −0.18 to +0.28) and presentation +0.28 (+0.10 to +0.47), on the 1 to 5 scale.
The study was sponsor-conducted: the analysis was executed by sponsor personnel against the
published plan and frozen before unblinding, and the raters, who scored blinded to summary origin, are
co-authors of the report. Two findings are relevant here. Inter-rater agreement was low across the
dimensions (Krippendorff’s α close to zero or negative for most elements; −0.26 for accuracy), and the
between-rater variance component was the dominant source of variation, larger than the case and caseby-arm components. And accuracy was the one dimension on which the primary Bayesian analysis did
not conclude non-inferiority (posterior mean −0.12, 95% credible interval −0.52 to +0.28; the point
estimate favoured the manual arm), although the pre-specified frequentist confirmation analysis did, so
the two models disagreed on that dimension alone; it was also the dimension with the lowest agreement,
the widest posterior and the largest dispersion. We draw two lessons from these results: 1) when a panel
of qualified clinicians is the comparator, the panel’s own agreement limits what can be concluded, and in
this study it was the limiting factor; and 2) rubric dimensions adapted from a validated documentation
instrument still left these raters without a shared standard, so calibration belongs before confirmatory use.
Both should be reported with the result. We suggest FDA expect inter-rater agreement to be reported for
any panel-based comparator, and expect rubric calibration to precede the confirmatory study. The study
evaluated summary quality under retrospective blinded assessment; it did not evaluate variantinterpretation software, clinical validity, or the clinical effectiveness of review in use.

6. R.2 and a limitation of our own architecture
R.2 asks whether properties hold for populations underrepresented in training data (p. 26). In genomics
this is the most important element in Appendix A, and it is a weakness of retrieval-grounded architectures.

The genomic evidence base is weighted towards European-ancestry populations [2]. A system that
faithfully retrieves it reproduces its skew. Traceability does not correct this and may make it worse: a real
citation can make an ancestry-limited assertion look better grounded than it is.
We propose two expectations. First, that the population composition of the evidence retrieved for an
output be a reportable property of that output and a reportable benchmark result, so that an inherited
limitation is visible instead of silently propagated. Where source metadata do not support this today, the
output should state that composition is not reported, rather than infer it; the gap belongs to the evidence
sources the class depends on, and FDA could name it. Second, that the system state explicitly, on the
output, when the retrieved evidence is thin or absent for the population the case concerns, where that
population is known to the system. R.2 for evidence-retrieval functions should cover the
representativeness of the evidence drawn on, and not only performance across user or patient subgroups
(Q9 asks about missing elements; this is one example).

7. Section V.B.2: rubrics, adjudicators, scope
Rubrics. Where multiple recognised frameworks apply, as they do for somatic variants, a sponsor should
state which framework and version a rubric implements and justify the choice.

Adjudicators. We support the principle that adjudicators be structurally independent of the sponsor and
the model developer, including when the adjudicator is an LLM (p. 14). Independent molecular-pathologist
adjudication at the volumes Q12 implies is a recurring cost that falls hardest on small developers. We ask
that shared adjudication capacity be part of the Q16 design.

Scope. The Discussion Paper indicates that benchmarking should have sufficient scope to cover clinical
domains relevant to the intended use, with comorbidity and drug-interaction examples (p. 14). There are
many and varied adjacent domains for genomic evidence. A bounding principle tied to the function’s
actual outputs would make the scope able to be estimated in advance.

8. Q19, Q22 and Q23: the evidence corpus as a change-control object
Section VI.C names three kinds of change: sponsor-initiated, model evolution, and third-party foundation
model (p. 20). The Discussion Paper also lists “retrieval strategies” as a change vector (p. 2), “other
components of the deployment architecture” as a re-benchmarking trigger (p. 19), and changes in “the
data environment” as a source of degradation (p. 20). None of the three named kinds covers, and none of
the other passages develops, the change that is characteristic of software of our kind: a change in the
content of an externally maintained evidence corpus in which an expert panel revises an assertion; a
framework issues a new version; a knowledgebase changes an evidence level. The sponsor did nothing,
the model did not change, the code did not change, but the output is appropriately different. Q22 asks
about categories of change; we propose this for consideration.

FDA has addressed adjacent questions. The 2018 guidance on the use of public human genetic variant
databases to support clinical validity, and the associated recognition programme, under which, to our
understanding, specified ClinGen expert-panel germline assertions and a specified portion of OncoKB’s
somatic content have been recognised, set expectations for sponsor reliance on an externally curated
resource, including curation quality, transparency and version control. Recognition is scope-limited: it
attaches to specified assertions, not to every entry or use, and says nothing about the quality of a
downstream system’s retrieval, synthesis or presentation. Within those limits it is the right foundation, and
we recommend FDA consider the two frameworks together.

We propose:
• Version pinning. The device should operate against, and disclose on the output, an identified
version of each source, so that any output can be audited against the evidence available to it.
 • Impact-tiered response to corpus change, as the trigger for corpus-driven reassessment (Q19
asks how triggering events should be determined), in place of an unqualified version-change
trigger, which would fire constantly, and alongside the periodic monitoring the Discussion Paper
describes for other sources of drift. Tier 1, no change to evidence content (metadata, synonyms,
formatting, citation corrections): document and monitor. Tier 2, a content change confined to
identified elements and expected not to alter any substantive evidence summary, uncertainty
presentation, escalation or output category: targeted regression against the affected benchmark
elements, and promotion to Tier 3 if the regression shows an effect that was not expected. Tier 3,
any change expected to alter one of those, and any broad change to a framework version,
ontology, retrieval ranking or source-inclusion policy regardless of expected effect: rebenchmarking of the affected functions against pre-specified thresholds, with risk review before
release. The auditable object is the sponsor’s tier-assignment procedure, with worked examples,
since the corpus changes themselves cannot be pre-specified (Q23).

9. Q21: responsibility and latency with an external curator
When an external curator revises an assertion, the curator has acted correctly. The failure that needs an
accountable party is the interval between the revision and the device’s propagation of it, and the handling
of outputs already released. We suggest: the curator is accountable for the assertion; the manufacturer is
accountable for its selection and use of the source, for detecting the change, propagating it, and running
a defined procedure for outputs released under the superseded state. The clock needs defined events. A
lookup timestamp shows what was fetched; it does not show how quickly a change became available
through the sponsor’s access, was detected, triaged, validated and deployed, or when affected prior
outputs were reviewed. We recommend risk-based, pre-specified and measured intervals for each
segment, scaled by the impact tier above, as a postmarket performance parameter for this class. Faster is
not always better: a validated propagation may properly take longer than an unassessed one, which is
why the interval should be pre-specified rather than minimised.

10. Q16 and Q26: third parties and retrieved content
We support third-party roles in maintaining sequestered challenge sets and providing shared adjudication
capacity, for the reasons in Section 4, and we accept the independence principles on p. 14 as the price of
a credible result. Q16 asks how to prevent third-party participation from limiting competition. Accredited
certification should not become the only route: sponsor-conducted benchmarking should remain an
acceptable route where the protocol is pre-specified, adjudicators are independent, development and test
material are separated, both verifiable from records on inspection, and protocol, criteria and results are
disclosed in full, including results that missed. Pre-specification and disclosure partly address Q10’s
concern about optimisation to the test; where benchmark-specific optimisation is a material risk,
sequestered or independently refreshed test material should be expected in addition.

On Q26, A.1 considers resistance to prompt injection “through user inputs, retrieved content, and tool
outputs” (p. 26). Retrieved content is an attack surface, and a retrieval-grounded system has that surface
whether or not it is agentic. Whatever boundary FDA draws for agentic systems, three expectations
should attach to retrieval-grounded devices: all retrieved content is treated as untrusted input, including
content from reputable curated sources; tool and retrieval outputs are checked against expected structure
before use; and no consequential action with effect outside the session, such as release of a report or
write-back to a laboratory system, occurs without explicit user confirmation. We would accept all three as
expectations on software of our kind.

11. Recommendations
1. Include traceability in the risk framework as claim-level measurements (attribution coverage,
attribution fidelity under blinded independent adjudication, critical omission, corpus version
disclosure, reconstructability), with no credit where fidelity is unreported or below pre-specified
criteria, the setting affords no verification, or the user cannot evaluate the sources (Q1, Q2).
 2. Treat the elevated consequence position for measurement-derived functions as rebuttable by
device-level evidence, with the degree of independent review demonstrated in use as the
evidence that counts (pp. 7-8, Q4).
3. For retrieval-grounded genomic functions, evaluate three endpoint families separately: evidence
fidelity; proposed-category performance against independent adjudication under a stated
framework and version, including whether insufficient or conflicting evidence is flagged; and the
reviewed workflow. Where the label is available to the system, treat agreement with it as
evidence of extraction fidelity, and require a blinded independent adjudicator where no
recognised assertion exists, sources conflict, or the draft extrapolates beyond the assertion’s
context (Q10, Q14).
4. Treat temporal holdout as insufficient rather than invalid: a supplement to evaluation of the
configured device against a pinned corpus snapshot, rather than a substitute for it (Q10).
5. Adopt a layered evaluation expectation: sponsor evaluation for overall performance and stress
tests, independent adjudication for fidelity and category endpoints, shared sequestered challenge
sets for stratified estimates; where stratified estimates cannot be obtained, state the limitation in
labelling (Q12, Q16).
6. Where a sponsor relies on human review, expect evidence that the configured workflow supports
it; expect inter-rater agreement to be reported for any panel comparator, and rubric calibration
before the confirmatory study (V.A, E.4, Q14).
7. Illustrate the Discussion Paper’s tailoring of confirmation approaches with a laboratory-workflow
example: retrospective blinded adjudication and prospective observation of the reviewed workflow
answer different questions and neither replaces the other; shadow deployment does not answer
the second (Q11).
8. Extend R.2 for evidence-retrieval functions to the representativeness of the evidence drawn on,
with population composition as a reportable property (or an explicit statement that it is not
reported) and a coverage warning on the output where the case population is known (Q9, R.2).
9. Add change in the content of an externally maintained corpus as a fourth change category, with
version pinning and a three-tier impact classification, considered alongside the public variant
database framework; treat detection-and-propagation intervals as a pre-specified and measured
postmarket parameter with defined start and stop events (Q19, Q21 to Q23).
10. Keep sponsor-conducted benchmarking an acceptable route under pre-specification, independent
adjudication, development/test separation and full disclosure, with sequestered test material
where optimisation to the test is a material risk (Q10, Q16).
11. Attach retrieved-content safeguards (untrusted-input handling, structural checks, confirmation
before consequential external action) to retrieval-grounded devices regardless of the agentic
boundary (Q26).

We would welcome the opportunity to discuss any of the above or to contribute to a workshop on this
subject.

Respectfully submitted,

James Creeden MD PhD
Chief Medical Officer
Qnomx AG, Basel, Switzerland
Switzerland Innovation Park Basel Area, Lichtstrasse 35, 4056 Basel, Switzerland
[1] Creeden J, Olivecrona M, Benavent N, et al. Evaluation of AI-assisted summarisation of tertiary clinical
genomics reports: results of the QNOMX-VHIR-CPSP-001 Phase 1 study. medRxiv 2026.
doi:10.64898/2026.09.10.26362656. Methods and statistical analysis plan: Creeden J, Olivecrona M,
Soriano A. medRxiv 2026. doi:10.64898/2026.06.11.26355467.
[2] Sirugo G, Williams SM, Tishkoff SA. The missing diversity in human genetic studies. Cell.
2019;177(1):26-31. doi:10.1016/j.cell.2019.02.048.