Limitations, assumptions, and future directions

Decision curves are modelled, not reconstructed

Moderate impact

Each model's published AUC is mapped to a sensitivity-specificity trade-off via a binormal approximation: assume normally distributed risk scores in cases and controls with equal variance and mean separation d' = √2 × Φ⁻¹(AUC). The binormal assumption is conservative for discrimination but loses information about the shape of the risk distribution (skewness, ceiling effects, tail behaviour).

What this means

The ranking of models by net benefit is robust. The exactnet-benefit values are not, especially at very low or very high thresholds. The clinically useful range's outer edges are approximations.

Workforce density is partly empirical, partly modelled

High impact

Real numbers: Queensland SA4s (Lindsay 2026, Australasian Journal of Dermatology, Early View) and New South Wales SA4s (Blake 2023). Modelled estimates: every other state uses Modified Monash Model (MMM) inferred density - we used the MMM remoteness classification to estimate workforce density for SA4s where no published count exists.

What this means

Victoria, Western Australia, South Australia, Tasmania, Northern Territory, and the ACT capacity numbers are MMM-anchored estimates, not measured. They should be read as order-of-magnitude. The QLD and NSW findings are the most defensible. The live map shows a provenance pill on every popup so the user always knows which is which.

The screening demand model is calibrated, not fitted

Moderate impact

We model the demand a given threshold creates as: of the eligible adult population (40-74, ≈48% of total ERP per AIHW), the fraction above threshold t is drawn from a log-normal risk distribution with mean equal to the national prevalence shifted by a regional incidence multiplier, and σ ≈ 0.8 in log-space.

σ ≈ 0.8 was calibrated so ~20% of eligible adults fall above the 3% threshold under national prevalence, matching the QSkin and 45 and Up risk-decile distributions reported by Olsen 2018 and Vuong 2016. It is not derived from individual-level data in this tool.

What this means

The qualitative pattern (high-incidence regions push more people above any threshold) is robust. The exact pop-above-threshold counts inherit the log-normal assumption and could be 10-30% off in either direction.

The workforce capacity model uses round-number constants

Moderate impact

Constant	Value	Source
Annual consults per dermatologist FTE	2,800	Lindsay 2026 central scenario
Annual consults per skin-cancer GP FTE	1,200	SCCA training program data
Screening fraction of clinical time	30%	Carved out from reactive workload
Screening cadence	2 years	BreastScreen, NLCSP, MSAC 1699
Risk horizon	10 years	Australian risk tools predict 10-year

What this means

These are central anchors with real-world spread. A reviewer should mentally test ±20% on each constant to see how sensitive a given finding is. The 30% screening-fraction anchor is the largest single point of leverage.

Modelled-extension workforce scenarios are speculative

High impact

Two of the five workforce scenarios in the picker are modelled extensions, flagged with an amber ~m badge:

Melanographer / Total Body Photography (TBP). We multiply effective derm throughput by 2.0× on the rationale that TBP decouples image capture (technician, high throughput) from image reading (specialist, async, batched). Defensible from first principles but awaiting ACEMID empirical data.
Tiered hub-and-spoke.GPs triage ~70% of patients locally at 1.5× normal throughput; specialists confirm ~30% at 0.5× normal load. Models the Roadmap's “tiered workforce” recommendation. Multipliers are nominal placeholders.

What this means

Numbers for these two scenarios depend on multipliers that have not been empirically validated. They illustrate the shape of what these workforce models could deliver, not the magnitude.

The 3%/year workforce growth rate is a central anchor

Moderate impact

The time-projection slider scales capacity by (1 + 0.03)^years. Defensible reasoning:

ACD pipeline: ~30 derm trainees/yr against a base of ~600 working derms = ~5%/yr gross, ~3%/yr net of retirement.
SCCA-trained GP pipeline is faster but smaller in absolute terms; weighted average pulls towards the derm number.
Roadmap Evidence Synthesis cites Lindsay 2026's central scenario at 2.5-3.5%/yr depending on state.

What this means

Sensitivity: 2%/yr → 1.22× over a decade; 4%/yr → 1.48×. The slider shows the active multiplier so reviewers can mentally swap.

Overdiagnosis adjustment inherits literature uncertainty

High impact

Named reference points in the OD slider: Lindsay 2024 in-situ (73.5% / 68.5%); thin-invasive 28% (range 22-34%); Glasziou 2020 all-melanoma women 54%, men 58%; Bjørch 2024 international range; Adamson 2024 US estimate.

We deflate net benefit by sens × prevalence × OD-rate × 1/(1-t). This is the standard Vickers-Elkin extension but inherits all of the OD-estimation uncertainty in the underlying literature, which is non-trivial: estimates vary widely by definition (in-situ vs invasive, thin vs thick), country, and ascertainment method.

What this means

The choice of OD rate dominates the deflation. A reviewer should always look at the chart at multiple OD anchors, not just one.

Incidence and population are static snapshots

Low impact

Incidence per SA4: AIHW Cancer Data in Australia 2024 + Cancer Atlas 2.0. Population: ABS Estimated Resident Population by SA4. Eligible adults: assumed constant 48% of total ERP.

No time-trend on incidence; the eligible-adult fraction doesn't age forward when the projection slider moves.

What this means

Reasonable for short projection horizons (≤5 yr). For 10-yr projections, ageing of the eligible population and any incidence trend should be modelled explicitly.

National prevalence is a hard-coded anchor

Low impact

We use 2.0% 10-year melanoma prevalence as the national default. The prevalence slider in the DCA section lets a user dial this between 0.5% and 5% (e.g. younger or older cohort), but the sidebar's regional aggregations always use the headline figure.

Equity stratification is descriptive, not causal

Moderate impact

The Equity Reveal section stratifies the per-region minimum feasible threshold by SEIFA quintile (ABS Index of Relative Socio-Economic Disadvantage) or ARIA+ remoteness (via MM1-MM7). The 1.5× disparity we surface (Q1 median min threshold 2.8% vs Q5 1.9%) is descriptive: it shows that the workforce maldistribution correlates with disadvantage.

What this means

It does notestablish that ATOF's adaptive threshold is the only (or best) policy lever to address the inequity. Workforce redistribution, alternative delivery models, and demand-side interventions are all complementary policy options.

The tool is not a microsimulation

High impact

ATOF is a decision-analytic dashboard, not an agent-based or microsimulation model. It does not:

track individual patients over time
model uptake / adherence heterogeneity
account for repeat-screening dynamics
model recall-rate or biopsy-cascade effects beyond the OD adjustment
carry uncertainty through the calculations (no Monte Carlo)
price cost-effectiveness directly (no ICER computation)

What this means

McLoughlin et al. (PLoS One 2025) is the microsimulation arm; ATOF is complementary. The two together cover both feasibility (microsim) and operationalisation (ATOF).

The risk-prediction models themselves have known weaknesses

Moderate impact

QSkin MP16 (AUC 0.76) is QLD-derived; external validity to non-Australian populations is untested.
MIA/Vuong (AUC 0.70) is the reference for 45-and-Up validation; parsimonious enough to deploy but trades discrimination for simplicity.
NCI MBRAT (AUC 0.70) is US-derived; ascertainment differs.
Cho 2005 (AUC 0.66) is the oldest and lowest-AUC model included; kept for historical comparison.

What this means

All AUCs are point estimates without confidence intervals propagated through ATOF's calculations.

Technical caveats

Low impact

Desktop-only. Best viewed on screens ≥ 1280px wide. A small-screen breakpoint shows a static summary instead.
Provenance pills only on the main map. The four equity-geography small-multiples do not show individual provenance. Reviewers should mentally flag any value from a non-NSW non-QLD SA4 as MMM-modelled.
Performance on lower-end machines. The 4-up small-multiples render ~350 SVG paths simultaneously and can be slow on older hardware.
Cookie/analytics consent is deferred until after the tour dismisses on first visit. A fresh visitor may not see the consent banner until ~1.5s in.

Future directions

The data sources, collaborations, and modelling extensions that would advance ATOF from the current draft to a more comprehensive analytic framework.

Individual-level 45 and Up Study sensitivity-specificity reconstruction

Retire the binormal approximation entirely. Reconstruct each model's sensitivity-specificity curve directly from individual-level 45 and Up risk scores and outcomes (Reyes-Marcelino et al., forthcoming via the Cust lab).

Outcome:the decision curves become citation-ready and the “modelled from published AUCs” caveat retires.

State-level workforce data for VIC, WA, SA, TAS, NT, ACT

Replicate the Lindsay 2026 supply-demand methodology for the six states currently relying on MMM-modelled density estimates. Likely partners: Australasian College of Dermatologists (ACD) state chapters, SCCA, AIHW Health Workforce data.

Outcome:every region's capacity number becomes empirical. Provenance pills go from mixed to uniformly “real.”

ACEMID empirical multipliers for TBP/melanographer scenario

The 2.0× derm-throughput multiplier for the Melanographer scenario is a first-principles defence. Empirical multipliers from the Australian Centre of Excellence in Melanoma Imaging and Diagnosis (ACEMID) would replace the placeholder.

Outcome:the “~m” modelled-extension badge can come off the TBP scenario.

Lindsay 2026 yearly workforce trajectories per state

The current time-projection slider uses a uniform 3%/yr compound growth nationwide. Lindsay 2026's underlying projection model produces state-by-state trajectories. Reading those values back into ATOF would let the slider track actual planned workforce growth.

Outcome: projection numbers stop being a global anchor and become spatially specific.

Confidence intervals throughout the sidebar

Carry uncertainty (Monte Carlo or analytic) from AUC CIs, OD-rate ranges, and capacity-constant sensitivity through to each sidebar metric. Surface intervals next to point estimates.

Outcome:reviewers can read the tool's claims with statistical context, not just point estimates.

Cost-effectiveness: from the beta calculator to a full ICER

A first step already exists: the Composite Calculator (beta) lets users enter their own workforce and cost assumptions and returns operational net benefit, total program cost, and cost per melanoma detected. It is honest about being an activity-cost model, not a cost-per-QALY.

The full version would add survival modelling to convert detected melanomas into quality-adjusted life years, then compute a true incremental cost-effectiveness ratio against the MSAC 1699 anchor (AUD $62,754/QALY) for any threshold and workforce scenario.

Outcome: ATOF becomes a useful input to HTA submissions, not just programme design.

Uptake and adherence modelling

ATOF currently assumes that everyone above threshold attends screening. Realistic uptake (BreastScreen 54%, NLCSP projected 60-70%) materially changes the demand side. Adding an uptake slider and adherence model would let programme designers compute realistic versus theoretical demand.

Outcome: the demand calculations become operationally meaningful, not just theoretical.

Microsimulation hand-off

A formal hand-off interface to the McLoughlin et al. PLoS One 2025 microsimulation. ATOF would surface programme-level scenarios; the microsim would simulate patient-level outcomes under each.

Outcome: a complete programme-design loop: pick a scenario in ATOF → simulate outcomes in microsim → feed cost back to ATOF for HTA.

Equity overlays beyond SEIFA and ARIA+

Add Indigenous status (AIHW Indigenous Health metrics), Modified Monash (DPA mapping), language other than English (ABS), and disability prevalence. Stratify the equity reveal on each.

Outcome: the equity conversation broadens beyond socioeconomic disadvantage to include structural and cultural dimensions.

Peer review and external validation

Submit a methods paper documenting ATOF's decision framework, validate the outputs against an independent cohort (45-and-Up + QSkin), and publish the codebase under an open license.

Outcome: ATOF becomes a citable tool, not a private dashboard.