ARS and ARDs in Action

R in Pharma European Summit 2026

Introductions

Orla Doyle 🇮🇪

Role: Executive Director
Head of Scientific Product Engineering
AQS
Organisation: Novartis
Focus area: Develops workflows and products for clinical trial reporting
Based in: Dublin, Ireland

Gregory Chen 🇨🇭

Role: Principal Statistician (Director)
BARDS HTA Statistics
Organisation: MSD
Focus area: Integrated evidence planning and synthesis via statistical techniques for HTA/value proposition/market access
Based in: Basel, Switzerland

Demystifying ARS

How we got here

ARS?

New idea
Let’s use ARS

A wall of interconnected notes

ARS is complex

Building a custom approach

Build our own

Regret after building our own

Despair

CDISC Analysis Results Standard logo

ARS v1.0 adoption

A familiar friend in clinical trial reporting - the adverse events table

Treatment-emergent adverse events by primary system organ class and preferred term

Adverse Event	Treatment A (N=xx)	Treatment B (N=xx)	Total (N=xx)
Number of subjects with at least one event	xx (xx.x)	xx (xx.x)	xx (xx.x)
Cardiac disorders	xx (xx.x)	xx (xx.x)	xx (xx.x)
Sudden death	xx (xx.x)	xx (xx.x)	xx (xx.x)
General disorders and administration site conditions	xx (xx.x)	xx (xx.x)	xx (xx.x)
Sudden death	xx (xx.x)	xx (xx.x)	xx (xx.x)
Postoperative fever	xx (xx.x)	xx (xx.x)	xx (xx.x)

Generating the Table: Parameters Are Embedded

Typical implementation — key parameters hardcoded directly in the script:

library(tidyverse)

# Safety population

adsl_safe <- adsl %>% 
  filter(SAFFL == "Y")              # <- hardcoded

n_trt <- count(adsl_safe, TRTA)    # <- hardcoded

# Treatment-emergent AEs
adae_te <- adae %>%
  filter(
    USUBJID %in% adsl_safe$USUBJID,
    TRTEMFL == "Y"                  # <- hardcoded
  )

# Count by System Organ Class and Preferred Term
result <- adae_te %>%
  group_by(AESOC, AEDECOD, TRTA) %>%  # <- hardcoded
  summarise(n = n_distinct(USUBJID), .groups = "drop") %>%
  left_join(n_trt, by = "TRTA") %>%
  mutate(
    pct  = round(n.x / n.y * 100, 1),
    cell = sprintf("%d (%.1f)", n.x, pct)
  )

Every filter, variable name, and hierarchy is buried in the code.
Changing the population means hunting through the script.

Separating Parameters from Code

Pull parameters into a YAML metadata file — the code becomes generic:

analysis_config.yaml

population:
  dataset:  ADSL
  variable: SAFFL
  value:    "Y"

events:
  dataset:  ADAE
  variable: TRTEMFL
  value:    "Y"

grouping:
  variable: TRTA

subject_id: USUBJID

hierarchy:
  - AESOC
  - AEDECOD

R code — driven by metadata

library(tidyverse)
library(yaml)

cfg <- read_yaml("analysis_config.yaml")

# Population — from metadata
adsl_safe <- adsl %>%
  filter(.data[[cfg$population$variable]] ==
           cfg$population$value)
n_trt <- count(adsl_safe,
               .data[[cfg$grouping$variable]])

# Events — from metadata
adae_te <- adae %>%
  filter(
    USUBJID %in% adsl_safe$USUBJID,
    .data[[cfg$events$variable]] ==
      cfg$events$value
  )

# Hierarchy — from metadata
result <- adae_te %>%
  group_by(across(all_of(
    c(cfg$hierarchy, cfg$grouping$variable)
  ))) %>%
  summarise(
    n = n_distinct(.data[[cfg$subject_id]]),
    .groups = "drop"
  )

Metadata Quickly Grows

One AE table means many analyses. The YAML expands rapidly:

reporting_event:
  id: "RE-AE-001"
  label: "AE Summary Table"

  analysis_sets:
    - id: AS001
      label: "Safety Analysis Set"
      condition:
        dataset:    ADSL
        variable:   SAFFL
        comparator: EQ
        value:      "Y"

  data_subsets:
    - id: DS001
      label: "Treatment-Emergent AEs"
      condition:
        variable: TRTEMFL
        value:    "Y"
    - id: DS002
      label: "Serious TEAEs"
      conditions:
        - {variable: TRTEMFL, value: "Y"}
        - {variable: AESER,   value: "Y"}
    - id: DS003
      label: "TEAEs leading to D/C"
      conditions:
        - {variable: TRTEMFL, value: "Y"}
        - {variable: AEACN,   value: "WITHDRAWN"}

  groupings:
    - id: GF001
      groupingVariable: TRTA
      dataDriven: false
      groups:
        - {id: G1, label: "Treatment A"}
        - {id: G2, label: "Treatment B"}
        - {id: G3, label: "Total"}

  methods:
    - id: M001
      label: "n and percent"
      operations:
        - {id: OP1, label: "n",
           function: COUNT_DISTINCT}
        - {id: OP2, label: "%",
           function: PERCENT_BY_SUBSET}

  analyses:
    - {id: AN001, label: ">=1 TEAE",
       dataset:     ADAE,
       analysisSet: AS001,
       dataSubset:  DS001,
       grouping:    GF001,
       method:      M001}
    - {id: AN002, label: "Serious TEAE",
       dataset:     ADAE,
       analysisSet: AS001,
       dataSubset:  DS002,
       grouping:    GF001,
       method:      M001}
    # ... 30+ more analyses per table
    # ... 100s of tables per study

Without a schema, this is unvalidated.
Without a standard, every team invents their own format.

Analysis Results Standard

ARS v1.0: Background

What it defines

A formal metadata model for clinical trial analysis results
Connects ADaM datasets (inputs) to results (outputs)
Covers concetps like populations, filters, groupings, methods, operations

How it helps

Unambiguous — no programmer interpretation required
Reproducible — any compatible tool gives the same result
Auditable — every result traces to a specification
Automatable — machines can execute the metadata directly

The ARS Stack

Layer	Purpose
LinkML	Tooling use to house the blueprint that defines and validates the ARS structure
ARS (before execution)	A machine-readable analysis specification
ARS (after execution)	The same specification, now enriched with results
ARD	The analysis results represented as structured result records that can be extracted from the completed ARS

The Wider ARS Landscape

The AE example is one slice of a broader results model: from study context and data inputs, through analysis definitions, to governed results and provenance.

Study and data context

Study and reporting event
Dataset and variable references
Controlled terminology
Analysis populations and subsets

Analysis definition

Analysis
AnalysisSet
DataSubset
GroupingFactor
Analysis method and operations

Results and hand-off

Result groups and values
Counts, percentages, summaries
Structured output records
Hand-off to tables, listings, and figures

Traceability and governance

Specification-to-result links
Versioning and provenance
Validation and conformance
Reproducible execution

The automation gap: ARS can make the result computable and traceable, but the final rendering and presentation metadata may live outside the model. Turning an ARD into a compliant table, listing, or figure still needs another governed layer.

This deck zooms in on the centre: how a familiar AE table becomes a structured, executable analysis definition.

Demystifying ARS, ARD, and LinkML

ARS = Recipe
The specification that describes what to analyze and how to compute it

ARD = Cooked Meal
The actual results produced by executing the ARS specification

LinkML = Recipe Template
Tooling to check the schema defining the structure and rules for writing valid recipes

Key Insight: ARS starts as a specification document. After execution, it can also contain the resulting ARD — the specification and its results travel together.

Worked Example: From AE Spec to Results

ARS Specification (Recipe)

analysis:
  id: AN001
  label: "Subjects with >=1 TEAE (%)"
  dataset: ADAE
  analysisVariable: USUBJID

analysisSet:
  condition: ADSL.SAFFL == "Y"

dataSubset:
  condition: ADAE.TRTEMFL == "Y"

grouping:
  variable: TRTA

method:
  operation: Operation_rollup_catvar_summ_grp_pct
  numerator: COUNT_DISTINCT(USUBJID) in dataSubset by TRTA
  denominator: COUNT_DISTINCT(USUBJID) in analysisSet by TRTA

↓ Execute

ARD (Cooked Meal)

analysis_id	analysisSet	dataSubset	group1_variable	group1_value	operation_id	raw_value	formatted_value
AN001	ADSL.SAFFL EQ Y	ADAE.TRTEMFL EQ Y	TRTA	Treatment A	Operation_rollup_catvar_summ_grp_pct	0.25	25.0
AN001	ADSL.SAFFL EQ Y	ADAE.TRTEMFL EQ Y	TRTA	Treatment B	Operation_rollup_catvar_summ_grp_pct	0.33	33.0

✓ Results

LinkML ensures both the specification and results follow the same schema — any tool that understands ARS can read and execute this.

ARS Model: Core classes for ARD generation

%%{init: {'theme':'default', 'flowchart': {'nodeSpacing': 24, 'rankSpacing': 32}, 'themeVariables': { 'fontSize':'12px', 'fontFamily':'arial'}}}%%
classDiagram
  direction LR

    class ReportingEvent {
    id
    label
    }
    class Analysis {
    id
    dataset
    analysisVariable
    }
    class AnalysisSet {
    condition
    }
    class DataSubset {
    condition
    }
    class GroupingFactor {
    groupingVariable
    }
    class AnalysisMethod {
    label
    }
    class Operation {
    label
    resultPattern
    }

    ReportingEvent --> Analysis : contains
    Analysis --> AnalysisSet : analysisSet
    Analysis --> DataSubset : dataSubset
    Analysis --> GroupingFactor : orderedGroupings
    Analysis --> AnalysisMethod : method
    AnalysisMethod --> Operation : operations

AnalysisSet = who (population)

DataSubset = which records (event filter)

GroupingFactor = how to split (columns)

Operation = what to compute

Mapping ARS Back to Our Table

Every element of the AE table maps to an ARS class:

Adverse Event dataset: ADAE AnalysisSet: SAFFL=“Y” DataSubset: TRTEMFL=“Y”	Treatment A (N=xx) GroupingFactor: TRTA	Treatment B (N=xx) GroupingFactor: TRTA	Total (N=xx) GroupingFactor: TRTA
Subjects with ≥1 TEAE, n (%)	COUNT xx PCT (xx.x)	COUNT xx PCT (xx.x)	COUNT xx PCT (xx.x)
Cardiac disorders	xx (xx.x)	xx (xx.x)	xx (xx.x)
Palpitations analysisVariable: USUBJID	xx (xx.x)	xx (xx.x)	xx (xx.x)
General disorders and administration site conditions	xx (xx.x)	xx (xx.x)	xx (xx.x)
Fatigue analysisVariable: USUBJID	xx (xx.x)	xx (xx.x)	xx (xx.x)

dataset ADaM source
AnalysisSet population
DataSubset row filter
GroupingFactor columns
analysisVariable unit counted
Operation computation

Demo: ARS-Driven Automation for CSRs

ARS and ARDs in Action: Market Access

Same Recipe, New Kitchen, New Diners

ARS: the recipe populations, subsets, groupings, methods, operations as data

ARD: the meal the result records an executed recipe yields

LinkML: the template what a valid recipe looks like

The CSR kitchen cooks for one diner: the regulator

One base ingredient, the trial data (SDTM, ADaM). One set menu, fixed by the sponsor in the SAP. One meal: the CSR tables and the integrated summaries (ISS, ISE).

Market access serves a table of HTA assessors

Each with their own tastes and allergies: a PICO (population, intervention, comparator, outcome), a template, and evidence they will not accept. Same base ingredient, plus ingredients from outside the trial, often not grown by the sponsor: published aggregates from a literature review, real-world data, registries, expert elicitation. The recipe has to adapt, not just re-run.

In the first published Joint Clinical Assessment, the HTA assessors could not reproduce most of the naive safety comparisons. The error was in an ingredient: event rates extracted from the comparator’s publication had been switched between two outcomes. It was caught because an HTA assessor tried to recompute the numbers from the dossier and could not. ^[1]

[1] Joint Clinical Assessment Report, Tovorafenib, v1.0, endorsed by the HTA Coordination Group 30 Apr 2026, published 9 Jun 2026, Appendix D.2, request 4 (p. 138): “The assessors were not able to reproduce the effect estimates shown for the naïve comparisons for most of the safety outcomes”; developer’s response: “the event rates for these two outcomes in the Bouffet 2023 study had been switched around in the analyses”. health.ec.europa.eu/publications/joint-clinical-assessment-report-tovorafenib-ojemda_en

Continuity: one strip, no re-definition. Orla’s stack table already reconciles the terminology: “ARD” is not an ARS class; results are OperationResult records inside the reporting event, and “ARD” is the flat dataset the R ecosystem extracts from them. Say “result records” when precision matters.
Why ARS exists, in one breath: twenty years ago we standardised what goes in (SDTM, ADaM, define.xml) and the document that comes out; the analysis in between stayed in code and its results in prose. ARS and ARD are the missing middle.
The reproduction story: tovorafenib JCA, request 4 of the second request for additional information. Serious AE, discontinuation due to AE and interruption due to AE could not be reproduced; severe AE could. The developer found the Bouffet 2023 event rates for two outcomes switched, corrected them and ran additional quality checks. Note for the discussion: the failure was in an external ingredient, not in the trial data, and it was found by hand.
Show of hands, 15 seconds: “Who has worked on an HTA dossier? Keep it up if you wrote code for one.” Calibrates depth, and tells you whom to call on in the discussion.

Two Asks of the Same Base Ingredient

%%{init: {'theme':'default', 'themeVariables': {'fontSize':'16px', 'fontFamily':'Source Sans Pro, arial'}}}%%
flowchart LR
  T["<b>Base ingredient</b><br/>trial data, one or more studies<br/>SDTM, ADaM"]
  X["<b>Added ingredients</b><br/>literature aggregates, RWD,<br/>registries, expert elicitation"]
  subgraph HTA["HTA track"]
    direction LR
    H1["<b>EU JCA</b><br/>relative clinical effectiveness<br/>and safety per PICO, with its<br/>certainty; no value judgement"] --> H2["<b>National HTA</b><br/>added benefit, cost-<br/>effectiveness, budget impact"] --> H3(["<b>Price, reimbursement,</b><br/><b>patient access</b>"])
  end
  subgraph REG["Regulatory track"]
    direction LR
    R1["<b>EMA</b><br/>quality, safety, efficacy;<br/>benefit–risk balance"] --> R2(["<b>Marketing</b><br/><b>authorisation</b>"])
  end
  T ~~~ X
  T --> R1
  X -. very occasionally .-> R1
  X --> H1
  T --> H1
  classDef reg fill:#eff6ff,stroke:#1e3a8a,color:#1a1a1a;
  classDef hta fill:#fef5e7,stroke:#e67e22,color:#1a1a1a;
  classDef ev fill:#f0fdf4,stroke:#16a34a,color:#1a1a1a;
  classDef ext fill:#faf5ff,stroke:#7c3aed,color:#1a1a1a;
  class R1,R2 reg;
  class H1,H2,H3 hta;
  class T ev;
  class X ext;

	Regulator asks	HTA body asks
The question	Do benefits outweigh risks for the indication?	How much better than what we use today, for whom, and is it worth the price?
Comparator, population	The comparator justified at design; the trial population	The local standard of care; the population the local pathway treats, often a subpopulation
Ingredients	Trial data, one or more studies; RWE the exception	Trial data, often several data cuts, re-analysed per template; plus SLR aggregates, real-world data, expert elicitation, costs and utilities: assembled by the sponsor, often not owned
Output	One body, one authorisation, one label	Many bodies, one dossier each: own populations, added-benefit category, cost per QALY, price and reimbursement

Regulatory track: once, EU-wide. HTA track: JCA in parallel with the EMA procedure, then national HTA, pricing and reimbursement country by country (EU mean 597 days from authorisation to availability, EFPIA W.A.I.T. 2025). Reg. (EU) 2021/2282, Recital 4, Art. 1(2), 2(6); Eichler 2010.

Purpose of the slide: enough context for the discussion, not an HTA course. The two tracks read the same base ingredient and ask different questions; the HTA question forces new analyses and new ingredients.
Regulatory: quality, safety and efficacy, with the benefit-risk balance weighing efficacy against safety (quality is a prerequisite, assessed alongside); the comparator is the one the sponsor justified at design, placebo or active. HTA: relative effectiveness against the comparator(s) that matter in each health system, plus national economic and organisational domains.
The Regulation’s Recital 4 names nine HTA domains: four clinical (health problem and current use, technical characteristics, relative safety, relative clinical effectiveness) and five non-clinical (cost and economic evaluation, ethical, organisational, social, legal). The JCA covers the four clinical domains only (Art. 2(6)); the five non-clinical domains and all pricing and reimbursement decisions stay national (Art. 1(2), Recital 14).
Timing: the JCA runs in parallel with the centralised procedure (dossier due at the latest 45 days before the envisaged CHMP opinion, Art. 10(1)); the report must be endorsed no later than 30 days after the Commission decision granting the authorisation (Art. 11(1)(a)) and is published after a procedural check; observed lag so far 40 to 69 days. National HTA and pricing follow the authorisation, and that is where most of the delay to access sits: EFPIA W.A.I.T. 2025 (published May 2026, medicines authorised 2021–2024) gives an EU mean of 597 days from authorisation to availability, median 532, ranging from 56 days in Germany to 1,201 in Romania.
Value proposition: in Germany the sponsor must demonstrate added benefit (Zusatznutzen) against the appropriate comparator therapy, on the basis of evidence it submits (§ 35a SGB V); in England the economic case is judged against NICE’s cost-per-QALY range, £25,000 to £35,000 since April 2026 (manual PMG36 §6.3). Both are claims the sponsor has to evidence, on top of the regulatory case.
Eichler et al. 2010 is the classic statement of the split: regulators license when benefits outweigh risks; payers decide on health benefit relative to existing treatment options.

The EU Joint Clinical Assessment in One Picture

%%{init: {'theme':'default', 'themeVariables': {'fontSize':'13px', 'fontFamily':'Source Sans Pro, arial'}}}%%
timeline
    title Regulation (EU) 2021/2282 roll-out
    Jan 2025 : New active substances for cancer
             : Advanced therapy medicinal products
    Jan 2028 : Orphan medicines
    Jan 2030 : All other centrally authorised medicines

How it runs ^[1]

Member States’ PICOs are consolidated into one scope with several PICOs; one JCA dossier is filed in parallel with the EMA application
A published report on relative effectiveness and safety, no value judgement, no price; national bodies “give due consideration”, then decide benefit, price and reimbursement
Commission tracker, extract of 10 Sep 2026: 23 started since January 2025, 4 published, 16 running, 3 discontinued (two for failing the Article 9 dossier requirements); published reports carried 1, 7, 8 and 12 PICOs ^[2]

Key clinical evidence in the dossier ^[3]

Per PICO: efficacy, safety, all subgroups with interaction tests, AEs by SOC and PT; each analysis flagged as prespecified or not, multiplicity-controlled or not, and who requested it
Where no head-to-head trial answers a PICO: indirect comparisons (Bucher, NMA, MAIC, STC) with their input data, model choices and convergence checks
CSRs, protocols and analysis plans as appendices; no individual patient data; code only “if the analyses and corresponding calculations cannot be described by a specific standard method”

[1] Regulation (EU) 2021/2282, Art. 7(2), 8, 9(1), 11, 13. [2] European Commission, JCA tracker (ongoing, completed, discontinued), data extracted 10 Sep 2026; published JCA reports: lurbinectedin (1 PICO), tarlatamab (7), tovorafenib (8), onasemnogene abeparvovec (12 in the final scope). [3] HTACG, guidance on the JCA dossier template for medicinal products (adopted 28 Nov 2024), App. A.2, A.3, B.2, D.3, D.5; guidance on multiplicity, subgroup and sensitivity analyses (10 Jun 2024).

JCA is joint but not binding: it replaces the clinical part of national assessments in spirit, not the national decision. “No value judgement” is Article 9(1).
The dossier template asks for interaction p-values for all subgroup analyses, full results where p < 0.05 or where a Member State asked, and all subgroup results in Appendix A.3. Adverse events by SOC/PT at ≥5% incidence in Appendix A.2, serious and severe regardless of incidence. The multiplicity guidance requires, for each analysis, whether it was prespecified, whether multiplicity was controlled, and who deemed an unplanned analysis necessary (sponsor, regulator, HTA body).
Indirect comparisons: dossier section 4.3.4 and Appendix D.5 (evidence-synthesis reports with protocols and SAPs); the reporting checklist in Appendix B.2 asks for software and version, code for non-standard methods, input data with derivation, MCMC chains, burn-in, iterations and convergence, the digitisation method for Kaplan-Meier curves, model and covariate selection, overlap and effective sample size, and results with and without population adjustment.
Appendix B.2: “the HTD is not required to provide individual patient data, only aggregated data such as effect estimates with standard errors are required”. Patient listings are supplied only on an Article 11(2) request and are not published. Appendix D.3 asks for code and output only for non-standard methods. So: documents and, conditionally, code. No datasets.
Tracker, 10 Sep 2026: 16 ongoing, 4 completed, 3 discontinued; catequentinib and the TIL therapy Tacquell were discontinued on 22 Jun 2026 because the dossier “fails to meet the requirements laid down in Article 9(2), (3) and (4)”; sasanlimab because the marketing application was withdrawn. Completed: tovorafenib (8 PICOs, only 1 assessable), tarlatamab (7), lurbinectedin (1), onasemnogene abeparvovec IT (12 PICOs across 6 populations in the final scope of 26 Feb 2026; several without comparative data).
Three of the four completed assessments rested on indirect comparisons: unanchored MAIC with an effective sample size of 5.82 (tovorafenib), Bayesian NMA plus MAIC on digitised curves (tarlatamab), anchored and unanchored MAIC plus ML-NMR (onasemnogene).
Article 13(1)(d): Member States may not re-request at national level what was submitted at EU level. HTACG Guiding Principles on Data Transparency (adopted 14 May 2026): trial results and evidence-synthesis methods are not commercially confidential; the dossier and its appendices are published, with redactions.

What HTA Analytics Look Like in Practice

Volume

German benefit dossiers grew from about 750 to 3,500 pages on average after the 2020 template change; outliers reach 20,000 ^[1]
Mandatory subgroups (sex, age, severity, region, stratification factors) for every endpoint ^[2]
Safety by SOC and PT with RR, CI and p, repeated per subgroup: one safety annex ran to 926 pages ^[3]
IQWiG sets aside 33% of AE analyses and 73% of subgroup categories submitted ^[4]

Ingredients that are not trial data: a meal rebuilt from its photograph

Comparators the trial never had: literature aggregates, RWD and registry controls, expert elicitation ^[5]
Survival curves digitised from a PDF and turned back into patients who never existed, then fed to NMA or population adjustment ^[6]
Three of the first four published JCAs rest on indirect comparisons; one on an effective sample size below six ^[7]
Thresholds are defaults, not a shield: 10 patients per subgroup, 10 events in one, interaction p < 0.05, incidence ≥ 5%. Filtered at 5%, tarlatamab’s JCA has no SOC breakdown of AE deaths ^[2,8,11]

Outputs that are not tables

Extrapolation parameters, transition probabilities and utilities feed a national cost-effectiveness model, outside the JCA’s clinical scope; NICE requires it fully executable, and its HTA assessors re-analyse it in 93 of 100 appraisals ^[9,10]
Countries file on different data cuts. Nobody requires an index of everything produced; the company needs one to know the totality of its own evidence
A dossier is read by HTA assessors who see at most patient listings, never analysis datasets they could re-run

[1] Schweitzer et al., Eur J Health Econ 2024;25(5): Modules 1–4 from 750 to 3,500 pages on average, “in individual cases 20,000 to even 40,000 pages”. [2] G-BA Module 4 dossier template, version 18 Nov 2025, §4.2.5.5, §4.3.1.3. [3] G-BA, empagliflozin Modul 4A Anhang 4-G, 28 Mar 2022, 926 pages; nintedanib Anhang 4-G, Feb 2025, 349 pages. [4] Schweitzer et al. 2024 (vfa, Dec 2022: 77% of submitted analyses not considered). [5] HTACG methodological guideline for quantitative evidence synthesis (8 Mar 2024); NICE DSU TSD 26 (expert elicitation). [6] Guyot et al., BMC Med Res Methodol 2012;12:9; HTACG practical guideline (8 Mar 2024). [7] Published JCA reports, Jun–Sep 2026 (tovorafenib ESS 5.82). [8] HTACG JCA dossier guidance (28 Nov 2024), App. A.2; IQWiG General Methods 8.0 §10.3.9. [9] NICE technology appraisal manual PMG36 §5.5.15. [10] Carroll et al., Value Health 2017;20(6): 100 single technology appraisals, exploratory analyses of the company’s model. [11] Tarlatamab JCA report, Table 98 note a.

The German numbers are the strongest case for “structure at scale”: the dossier grew 4 to 5 times after 2020, mostly AE and subgroup analyses, and IQWiG disregards a third of AE analyses and nearly three quarters of subgroup categories. Around €800k to €1m and 12 months per dossier (vfa 2022; Schweitzer 2024).
Macabeo et al. (PharmacoEconomics Open 2024;8(1)): 22% of 543 oncology HTA reports across five countries used an indirect comparison; 30% of those were accepted and about half were left unclear. The JCA experience so far is harsher: tovorafenib’s MAIC matched on an effective sample size of 5.82.
The middle column is where ARS, as designed, stops applying: the ingredients are not ADaM and the methods carry assumptions (effect modifiers, weights, priors) that a “method + operation” model does not capture. The photograph line is the one metaphor extension worth making: pseudo-IPD reconstruction really is recovering a dish from a picture of it.
Right column: the model is the consumer of many results, and the HTA assessor re-analyses the model without the analysis datasets (at most CSR patient listings). The data-cut point is about the company’s own sanity: a filing in one country may rest on interim analysis 2, in another on the final cut; efficient production and an accurate index are for the sponsor to know what it has produced, not because anyone mandates it.

The Analysis Space Is Combinatorial

Analysis families

Descriptive: baseline, disposition, drug use before, during, after
Efficacy, PRO, quality of life
Safety: overall, serious, grade ≥3, by SOC and PT
Subgroups, interaction tests, sensitivity
Model inputs: transition probabilities, PRO-mapped utilities

SAP and each HTA template

Populations

Label population
Subpopulations along the local pathway: line, biomarker, prior therapy, each with its own comparator

Member States’ PICOs, national pathways

Comparators and their ingredients

Trial control arm
Standard of care per country: indirect comparison on literature aggregates, digitised curves
External controls: RWD, registries
Expert elicitation where data are absent

the HTA assessor, not the sponsor

Data cut-offs

Regulatory cut
Later cuts for each region’s timing
Different countries file on different cuts

submission calendars

Templates

EU JCA dossier
G-BA Module 4
NICE, HAS, AIFA, …
Own effect measures, thresholds, layouts

each HTA body

= thousands of result records, each one analysis serving a PICO: population, intervention, comparator, outcome. ARS names the same ideas differently: analysis set, grouping, data subset, method. Same recipe structure, different vocabulary: aligning it, and extending it where HTA has words ARS does not, is the next two slides.

Message: the volume is not exotic analyses, it is ordinary analyses multiplied by factors the sponsor does not control. That is exactly the situation a specification-driven approach is meant for.
Populations and comparators are coupled: a subpopulation defined by the local pathway usually implies a different comparator, so a new PICO. Subgroups within a PICO do not create new PICOs (HTACG scoping guidance, Nov 2024).
Model inputs are analyses too: parametric survival fits for extrapolation, transition probabilities, utilities from mapping PRO instruments to EQ-5D with country value sets. They are produced from the same ADaM and rarely appear in a CSR.
The waste argument cuts both ways: IQWiG sets aside 73% of subgroup categories (Schweitzer 2024). Own it: automation makes the waste cheaper, not smaller. Navigability is the way out.
Ask the room: how much of your HTA programming is genuinely new analysis, versus the same analysis with a different population, comparator or threshold?

What ARS Names, and What HTA Would Add

ARS class	In the earlier AE table	In the HTA setting	What to record, for whom ^[1,2]
`ReportingEvent`	The CSR	One dossier per body, data cut and template version	Which body, cut and template: to find the right dossier, and the delta between two	fits
`AnalysisSet`, `DataSubset`	`SAFFL = “Y”`, `TRTEMFL = “Y”`	The PICO population, often a subpopulation of the label; event subsets unchanged	Which PICO the population answers: so the HTA assessor can find the analysis they asked for	fits
`GroupingFactor`	`TRTA`; SOC and PT, data-driven	Treatment vs the local comparator; effect-modifier subgroups	The comparator’s identity and where its data came from: trial arm, publication, registry, elicitation	stretch
`AnalysisMethod`, `Operation`	COUNT, PERCENT	RR with CI and p; weighted HR after matching; network meta-analysis	The assumptions behind the estimate: matched effect modifiers, weights and effective sample size, priors, convergence. The JCA checklist asks for exactly these	stretch
`Analysis.reason`, `purpose`	SPECIFIED IN SAP; PRIMARY OUTCOME MEASURE	Requested by a Member State; post hoc for PICO 5; requested by an HTA assessor	Prespecified or not, multiplicity-controlled or not, and who asked: required per analysis by the JCA guidance. ARS has four reason terms, none of them HTA	stretch
`OperationResult`, `ResultGroup`	n (%) per cell	RR, CI, p per PICO × subgroup × cut	Identifiers tying each number to its PICO, cut and template: for navigation, and for a model to cite a record instead of a page	fits
Programming code, `context`	The R script	Code per analysis, as JCA Appendix D.3 asks	The code and its environment: reproducibility for the sponsor, legibility for an HTA assessor who cannot run it	fits

[1] HTACG JCA dossier guidance (28 Nov 2024), App. B.2 reporting checklist for evidence syntheses: software and version, code for non-standard methods, input data and derivation, MCMC chains and convergence, digitisation method, model and covariate selection, overlap and effective sample size. [2] HTACG guidance on multiplicity, subgroup and sensitivity analyses (10 Jun 2024): for each analysis, prespecified or not, multiplicity control, and by whom it was deemed necessary. ARS v1.0 classes and terms checked against the published LinkML model (19 Apr 2024). “Fits”: covered as is. “Stretch”: a mechanism exists; the vocabulary does not.

Read the table left to right: Orla’s example, the HTA equivalent, and then the only question that matters for this session: what would need to be recorded, and who would read it.
The right-hand column is not invented: it is what the JCA dossier guidance already asks for in prose. The question is whether it lives in a Word appendix or in the result record.
“Stretch” rows: ARS has the mechanism (categories, reason codes, reference documents, a free-text dataset name) but no shared terminology for PICO, HTA body, comparator source or synthesis assumptions. Analysis.reason offers SPECIFIED IN PROTOCOL, SPECIFIED IN SAP, DATA DRIVEN and REQUESTED BY REGULATORY AGENCY, extensible through TerminologyExtension.
Terminology guard: “ARD” is the flat dataset extracted from OperationResult records; say “result records” when a CDISC person is in the room.

What ARS Has No Word For Yet

Criteria for execution

“PT only if incidence ≥ 5%.” “Subgroup curve only if the interaction is significant.” The rule tests a result, not a record, so no where-clause can express it. The template decides what is shown by default.

Record the rule, or only its outcome?

Ingredient provenance

A digitised curve, an event rate copied from a publication, an expert’s prior. ARS can name a source document; it cannot say how the ingredient was made. The first JCA’s error lived exactly here.

Record the recipe for the ingredient too?

Where a number goes

ARS knows a result’s origin and its display. A survival parameter that feeds a cost-effectiveness model has a destination, not a table cell.

Record where a number goes, or only where it came from?

Environment and change

One field for the language and version. Version numbers that say which, never what changed. A delta dossier is a diff nobody can compute.

Record the environment nobody re-runs? Record the diff?

Record as much as possible? How much is too much, and for whom? Each field on this slide costs a statistician time now; the benefit lands later, and often with someone else. That trade is the discussion.

Position of the slide: these four are the concepts ARS v1.0 has no class for, as opposed to the previous slide’s mechanisms that exist but lack vocabulary. Tooling for some of them exists outside ARS (the hta-pharma packages encode AMNOG criteria; survival packages hand parameters to models), which is why this is a discussion, not a gap list: nothing here is integrated into the standard.
Criteria: the G-BA and JCA rules (10 patients per subgroup and 10 events in one, interaction p < 0.05, incidence ≥ 5%) all test results, not records.
Provenance: the tarlatamab MAIC code reads a file literally called pseudo_ipd.csv. Nothing in the dossier’s metadata says how that file was made. The tovorafenib error was a switched event rate copied from a publication.
Environment: this row feeds Prompt 3 in the discussion. The version question feeds the delta-dossier reality: a JCA report may not exist when the German dossier is due (EFSPI 2025).
Hand over to the “who asks” slide: before deciding what to record, look at who is asking for anything at all.

Who Asks for Data and Code Today

Body	Asked for today	Direction of travel
FDA	SDTM, ADaM, define.xml required; source code for ADaM and primary and secondary efficacy outputs expected ^[1]	Dataset-JSON consulted in 2025, not yet adopted; the Data Standards Catalog lists Define-XML only, no ARM, no ARS ^[2]
EMA	Nothing required for a marketing application. Voluntary pilot since Sept 2022, 13 procedures: SDTM, ADaM, programs, ARM “highly recommended” ^[3]	Pilot extended “until further notice”; the network workplan plans a transition to systematic submission of clinical study data in 2026–2028, alongside the revised EU pharmaceutical legislation ^[4]
EU JCA	Documents per PICO; no individual patient data; code only for non-standard methods, and in practice as PDF ^[5]	AI principles for dossiers, July 2026: every AI-assisted step documented, prompts kept, a human responsible ^[6]
G-BA, NICE	G-BA: full CSRs, readable code when deviating from standard software, code plus input data for indirect comparisons. NICE: a fully executable economic model with full code access ^[7,8]	G-BA template (Nov 2025) cross-references the EU dossier; NICE position statement on AI in evidence generation (Aug 2024) ^[7,9]

Nobody asks for ARS: not one regulator, not one HTA assessor. And no HTA assessor has the ADaM. FDA can re-execute a sponsor’s analysis and EMA is heading there; no HTA body can, by design. So “reproducible” cannot mean at the G-BA what it means at the FDA. What is left is legibility: what was done, for whom, against what, at which cut.

[1] FDA Study Data Technical Conformance Guide v6.2.1 (Jun 2026) §4.1.2.10. [2] FDA Data Standards Catalog v11.0 (Mar 2025); Federal Register, 9 Apr 2025 (Dataset-JSON request for comments). [3] EMA, clinical study data pilot Q&A (EMA/658116/2022, rev. 20 Jan 2026) §3.3–3.4; NDSG annual report 2025 (Jan 2026). [4] EMA Network Data Steering Group workplan 2026–2028 (Feb 2026); EMA Programming Document 2026–28. [5] HTACG JCA dossier guidance (28 Nov 2024) App. B.2, D.3; published JCA dossiers Jun–Sep 2026. [6] HTACG, general principles on the use of AI in JCA dossiers (adopted 15 Jul 2026). [7] G-BA Module 4 template (18 Nov 2025) and dossier format document (8 Apr 2026). [8] NICE technology appraisal manual PMG36 §5.5.15 (updated 31 Mar 2026). [9] NICE ECD11 (15 Aug 2024).

FDA wording (Study Data Technical Conformance Guide v6.2.1, June 2026): “Sponsors should provide the source code used to create all ADaM datasets, tables, and figures associated with primary and secondary efficacy analyses.” The Data Standards Catalog v11.0 lists Define-XML; neither ARM nor ARS appears in the Catalog or the Guide.
EMA: no general requirement today; the clinical-study-data pilot uses CDISC SDTM/ADaM plus programs and “highly recommends” ARM; 13 procedures selected by December 2025 (NDSG 2025 annual report). The NDSG workplan (Feb 2026) says the pilot “will transition to systematic submission of clinical study data for CAPs” in 2026–2028, in light of the revised pharmaceutical legislation, expected to apply from around 2028. A stated plan, not yet a requirement.
PMDA, for completeness: SDTM and ADaM required; programs for key analyses required but need not be executable; ARM “preferable” in define. It is the regulator that pushes ARM hardest.
Appendix D.3 in the four published dossiers: one set of code removed before publication; one set of 29 print-to-PDF files with no extractable text; one R snippet with a sessionInfo(); SAS programs that read ADaM datasets no HTA assessor will see.
The asymmetry to name out loud: by 2028 the regulator may hold ADaM and programs for every centrally authorised product; the JCA runs in parallel on the same trial and receives PDFs. How long does that hold, and which side moves first?
HTACG AI principles (adopted 15 Jul 2026): developers stay accountable; no step “fully automated without a human being ultimately responsible”; document tool name, version, date, developer and purpose for every AI-assisted step; “all prompts used in AI tools are recorded and made available if requested”.
Nobody asks for renv lockfiles, containers or package manifests. That is Prompt 3 in the discussion.

Where This Leaves Us

Re-use is highest in HTA

One recipe, many diners, many cuts. A machine-readable recipe pays back on every re-cook, and nowhere is it re-cooked more.

A shared standard is what makes shared tooling possible

Thousands of records to navigate, trace and diff. Sponsors need that for QC, HTA assessors to find what they asked for. Nobody builds a viewer for an in-house format.

Nobody has asked, and every field has a cost

No regulator or HTA assessor names ARS. Each field costs a statistician time now; the benefit lands later, and often elsewhere: a later dossier, a partner, an HTA assessor.

What would we want to achieve, and what would it cost to capture? Reproducibility, efficiency to reproduce, or legibility for a reviewer and for a machine: pick the goal before picking the metadata. Over to the discussion.

Card 1 answers “why would HTA benefit more”: not because HTA analyses are harder, but because the same analysis is re-specified more times, by more parties, on more deadlines. Re-use is where a specification pays back.
Card 2: a standard is not only about reproducing a number, it is what lets a common navigator exist. Sponsors already need one internally to QC thousands of outputs; HTA assessors need one to find the analysis they asked for. Keep it as a question if challenged: would an HTA assessor use a viewer the sponsor built, and should they?
Card 3 pre-empts the expert’s question (“why a standard and not a good internal format?”) and names who pays: the cost falls on the sponsor’s statistician; the benefit on a future team, a partner or an HTA assessor. That externality is why voluntary standards stall.
Transition slide into the discussion section. Do not answer the question; hand it over.

Discussion and Q&A

From Structured Results to Trusted Evidence

ARS can make an analysis computable. What makes it usable, admissible, and trusted beyond the team that produced it?

01 · Efficiency ↔︎ readability

Can we accelerate production without making the result harder for a statistician, reviewer, or model to inspect?

02 · Submission ↔︎ evidence

Should a regulatory or HTA submission carry the ARS/ARD, the rendered output, the execution environment, or all three?

03 · Automation ↔︎ validation

Does machine-readable metadata reduce validation effort, move it earlier, or create a new object that must be validated?

04 · Results ↔︎ decisions

Which PICO, comparator, estimand, data cut, uncertainty, and presentation details must be machine-readable for an HTA assessor to challenge the result?

The question for the room: where should the standard stop, and what governed layer should come next?