Context

NSW Ambulance publishes annual OHCA summary statistics but does not release patient-level event data. Published research and aggregate Ambulance Service reports put median response times in Greater Sydney between eight and twelve minutes, well outside the window in which defibrillation changes survival odds for shockable rhythms. Bystander AED retrieval is rare in residential settings because the nearest registered device is often inside a closed commercial building.

The LIFT programme treats this gap as an engineering and logistics problem first, and a clinical problem second. If the defibrillator can be flown to the patient in under four minutes from a rooftop cradle, the chain of survival reconfigures: the bystander performs CPR, the drone delivers the AED, and the ambulance arrives to transport, not to resuscitate.

Scope

LIFT Study 01 is bounded to the Greater Sydney Statistical Area and to the DJI FlyCart 30 as the reference airframe. It does not attempt to model regulatory approval timelines, nor to specify any particular procurement pathway. It produces transparent, reproducible coverage maps and cost envelopes that a decision-maker can interrogate line by line.

Midnight rectangle with a signal-red rule and the words “hello world” set in the display typeface.

The scaffold build uses the following reference toolchain:

Stage	Tool	Version
Markdown parse	Pandoc	3.1.3
Print rendering	paged.js via Chrome	147.0

Survival depends on the chain. Break any link and the chance of survival collapses.

Resuscitation Council UK, 2025

The remainder of the report is written into this scaffold in later sprints. For now, the measure of success is narrow and specific: both public/index.html and public/lift-study-01-ohca-aed.pdf must build from this chapter without errors, and both must render the heading hierarchy, the figure, the table, and the pull quote above.

Chapter 02

Data visualisation

---

Sprint 20 introduces the charting pipeline that will carry every quantitative finding in LIFT Study 01. Three reference charts, generated from real project data, exercise the full pipeline end-to-end: CSV in, design-token-styled SVG out, same source consumed by both the web and PDF builds. Every chart below is regenerated from its source dataset by npm run build:charts; there is no manual editing stage.

Platform scoring

FlyCart 30 is highlighted in signal red as the top-scoring platform; the remaining four are rendered in slate grey for context.

Key insight. The DJI FlyCart 30 leads the feasibility shortlist with a weighted composite of 3.49, sixteen per cent above the runner-up (KITE, 3.02) and forty-six per cent above the lowest-scoring platform in the shortlist. The FlyCart’s advantage is concentrated in the two attributes that matter most for AED delivery missions: maximum payload (top score of 5.0, reflecting the 40 kg gross payload headroom) and field maintainability (5 out of 5). The trade-off is range — FlyCart scores only 0.46 on that axis — but for the suburban Sydney mission profile, a 4–8 km catchment radius makes payload, not range, the binding constraint. The ranking is provisional and carries the sprint02_v1_heuristic label; it should not be read as a procurement recommendation.

Coverage versus base count

The base scenario is drawn in signal red as the headline curve; conservative and optimistic are rendered in slate grey as sensitivity context.

Key insight. Marginal coverage collapses rapidly beyond the eighth base. In the 6 km base scenario, bases one through five deliver 39.9 pp of coverage (77 per cent of the total achievable with the evaluated candidate set), while bases nine and ten contribute exactly zero additional percentage points. The same saturation pattern appears in the conservative and optimistic scenarios, differing only in the absolute ceiling reached (43.9 per cent at 4 km; 51.4 per cent at 6 km; 64.7 per cent at 8 km). Operationally this says two things. First, the initial rollout design should target six to eight bases, not ten: the last two sites are coverage-redundant against this candidate list. Second, lifting the achievable ceiling past roughly two-thirds of synthetic hotspots requires either a larger catchment radius (driven by airframe range) or a richer candidate base list — not more sites from the current shortlist.

Distance versus time advantage

Each marker is an SA2; markers above the break-even line are drone-faster, markers below are ambulance-faster and are highlighted in signal red.

Key insight. The drone time advantage is strongly distance-bounded. Of the 37 SA2 areas in the comparison, 27 see a positive time advantage over the 8.5-minute ambulance baseline — but every one of the 10 negative-advantage areas lies beyond roughly 11 km from the nearest base, and the worst case (Penrith, 24.5 km) loses 8.2 minutes. The implicit break-even point sits near 11 km, consistent with the drone kinematic model (2 minutes on-base plus 0.8 minutes per kilometre). This has direct siting implications: the existing 10-base shortlist leaves the outer western corridor — Penrith, Mount Druitt, St Marys, Liverpool — structurally underserved, and no amount of re-sequencing the shortlist fixes it. A second base ring beyond the current BASE_010 cluster is needed before the drone network can claim system-wide parity, let alone advantage, across Greater Sydney.

Reproducibility

All three figures are deterministic outputs of the scripts/charts/build-*.js generators. The pipeline reads design tokens from design-system/tokens.json at runtime; no colour, font, or sizing value is hardcoded in chart code. Running npm run build:charts regenerates every SVG from its source CSV, and npm run build:all wires the charts into both the web build (public/index.html) and the print build (public/lift-study-01-ohca-aed.pdf).

A defibrillator-carrying drone network is, at heart, a spatial optimisation problem: where the events occur dictates where the rooftop cradles must be, and how far a drone must fly to reach each new collapse. This chapter characterises the spatial distribution of out-of-hospital cardiac arrest (OHCA) across 37 SA2s of Greater Sydney using the synthetic cohort carried in data/derived/synthetic_ohca_hotspots.csv. Two complementary views are shown below: a rate choropleth that normalises for population (task 3.3), and a proportional-symbol map that reports absolute event volume (task 3.4). Read together, they reveal a pattern that is not obvious from either map alone.

Synthetic-cohort caveat. The figures and rankings below descend from a proxy dataset that combines the ABS 2021 Greater Sydney SA2 population and age structure with the ABS 2021 SEIFA Index of Relative Socio-economic Disadvantage (IRSD), anchored to a Hasselqvist-Ax 2015 incidence base rate. The derivation is documented in full in Appendix M — Methodology: synthetic OHCA proxy with SEIFA IRSD weighting. Treat every absolute count on this page as pending verification against NSW Ambulance incident data; relative ordering between SA2s is expected to hold under replacement with the real cohort.

Where rates concentrate

The SA2 polygons sit on a muted Sydney base map (CartoDB Positron, © OpenStreetMap contributors) showing the coastline around Sydney Harbour, Botany Bay, the Parramatta and Georges rivers, and the principal road network — so readers can locate the hotspots against a recognisable geography. The data layer uses a ColorBrewer 5-class sequential-reds palette. Dark shading clusters across Sydney’s south-west corridor (Fairfield, Cabramatta, Bankstown, Campbelltown); the palest shading covers the northern shore and the outer growth LGAs such as Kellyville and Rouse Hill where the IRSD decile is high and the age-65+ share is low. A signal-red callout ring marks the highest-rate SA2. A quantile legend, scale bar, north arrow, base-map attribution line, and data-provenance footnote sit adjacent to the map.

Key insight. Rate concentration is a joint geography of age and disadvantage. Under the SEIFA-IRSD-weighted proxy (Appendix M), the highest synthetic rates sit in Sydney’s south-western corridor — Fairfield, Cabramatta–Lansvale, Bankstown, Campbelltown-South and Campbelltown-North all land in the top quantile. Age structure still matters, but the IRSD weighting pulls the rate surface decisively away from the leafy northern-shore band the earlier age-only proxy highlighted. The policy implication is inverted from that earlier reading: a “wealthy boomers” heuristic would mis-site base placements away from the SA2s the combined age-plus-deprivation signal actually prioritises. The choropleth also carries the full MAUP caveat: re-aggregating the same cohort to SA3 or LGA boundaries would smooth the south-west cluster into a single mid-range polygon, and pooling across LGA would re-introduce the wealthy-suburb age-signal artefact.

Where the volume sits

The 37 centroid circles sit on a muted Sydney base map (CartoDB Positron, © OpenStreetMap contributors) showing coastline and the principal road network so each symbol reads against a recognisable Sydney geography. Circle sizes range from small (tens of synthetic cases in wealthier northern SA2s) to large (more than 200 synthetic cases in the south-west cluster led by Fairfield, Bankstown and Cabramatta–Lansvale). Three signal-red callout rings label the top-ranked hotspots. A graduated-circle legend shows minimum, midpoint, and maximum case counts, with a Viridis class key underneath. Scale bar, north arrow, base-map attribution line, and a data-provenance footnote sit adjacent to the map.

Key insight. Volume and rate now point at the same part of the city — the disadvantaged south-west. The top-three SA2s by synthetic volume — Fairfield, Bankstown, and Cabramatta–Lansvale — between them carry approximately 16% of synthetic cases, and the top-five extends this cluster south to Campbelltown-South and Campbelltown-North. Every SA2 in the top-five sits in SEIFA IRSD decile 1 or 2 (most disadvantaged), confirming that the combined age-plus-IRSD proxy is the logical first-pass driver of drone-base placement. Note the deliberate encoding: a choropleth would misrepresent this view because event counts (as opposed to rates) are an extensive variable and must not be mapped by area fill. The proportional symbol map is the cartographically honest form for the question “where are the biggest hotspots, in absolute terms?”

Caveats and next steps

Both maps draw on the synthetic cohort only; neither represents observed NSW Ambulance incident data. From Sprint 26 onwards the synthetic proxy is SEIFA-IRSD-weighted; the derivation, the Hasselqvist-Ax 2015 base-rate anchor, and the pending-verification labelling convention are documented in Appendix M — Methodology: synthetic OHCA proxy with SEIFA IRSD weighting. The synthetic counts remain above the n < 5 suppression threshold for every SA2, so no cells are suppressed, but rates are not age-standardised and will need Empirical Bayes smoothing when the real cohort is slotted in. The nearest-centroid raster proxy used in Figure 3.1 is a presentational stand-in for ABS SA2 polygons and is explicitly labelled as such on the map itself. The final report build will swap in the true SA2 geometry and re-run the quantile classification; the class breaks will shift slightly but the south-west cluster is expected to survive the substitution because the ranks are carried by the underlying rates, not by the polygon shapes.

This chapter is a build-time test artefact. It exercises the full Sprint 22 citation pipeline — Pandoc citeproc plus pandoc-crossref — against the bibliography assembled in Sprint 21. Every claim below carries a real reference drawn from content/references.bib; no citation is fabricated, and no citation is typed manually. The chapter also exercises the cross-reference syntax so that Chapter sec. 4 references itself, Figure fig. 1 is numbered, and Table tbl. ¿tbl:citation-types? is numbered in both web and print outputs.

Clinical evidence base

The time-critical nature of out-of-hospital cardiac arrest is established in foundational work [1,2] and reinforced by later survival modelling [3,4]. Delayed defibrillation is the dominant modifiable cause of poor outcome even inside hospitals [5], and bystander CPR alone cannot substitute for early shock delivery [6]. Cost-effectiveness analyses of lay defibrillation programmes [7p. 227] found that community placement of AEDs pays back within conventional willingness-to- pay thresholds, though uptake in Australia remains uneven [8].

The Institute of Medicine’s strategic review of cardiac-arrest survival [9] named geographic and temporal access to AEDs as the single largest leverage point — a finding echoed by systematic analyses of public-access defibrillation failure modes [10].

Drone-delivered AED evidence

Drone-delivered defibrillation moved from simulation [11] to live prospective deployment in Sweden [12,13], and the 2023 observational series from the same group [14] reports drone arrival ahead of ambulance in roughly two-thirds of suspected OHCA events. Systematic-review coverage [15] synthesises the small but coherent literature, and recent Canadian modelling [16] concludes drone-delivered AEDs are cost-effective under realistic deployment densities.

National-Institutes-of-Health funding for cardiac-arrest research [17] lagged the disease burden for the decade preceding these trials, which is one reason the Swedish operational evidence carries disproportionate weight.

Methodology literature

Systematic review method is grounded in PRISMA 2020 [18] and the PRISMA-S search-reporting extension [19]. Grey-literature search follows Godin et al. [20], and the Bramer algorithm [21] guides the MEDLINE / Embase build. Qualitative synthesis of heterogeneous evidence draws on critical-interpretive synthesis [22] and meta-narrative review [23] — the latter matters here because the drone-AED literature crosses clinical, aeronautical-engineering, and logistics disciplines with incompatible vocabularies.

Grey literature and operational evidence

Australian incidence data are drawn from the AIHW annual OHCA series [24] and state-level ambulance reporting [25,26]. Population denominators come from the 2021 Census [27] and regulatory constraints from CASA’s Part 101 manual of standards [28]. International comparators include the OHCAO registry for England [29].

Clinical guidance is anchored in the 2020 American Heart Association resuscitation guidelines [30]. Operator experience in medical drone logistics is captured through company grey literature for Zipline’s Rwandan network [31], Everdrone’s Swedish deployments [32], Swoop Aero’s Malawi and Pacific operations [33], UNICEF’s humanitarian-health drone programme [34], and the NHS blood-sample trial [35].

The recent analysis of LLM-assisted biomedical writing [36] is cited here as a methodological caution for any AI-authored synthesis appearing in this report.

Citation-type coverage

Table tbl. ¿tbl:citation-types? summarises which source categories the bibliography covers and which single citation in this chapter exercises each type.

Section sec. 4.1 establishes the clinical baseline; Section sec. 4.2 surveys the drone-AED evidence; Section sec. 4.3 names the review methods; Section sec. 4.4 enumerates grey-literature and dataset sources. Figure fig. 1 provides the pipeline diagram referenced above.

References

This chapter documents the derivation of the synthetic out-of-hospital cardiac arrest (OHCA) counts carried in data/derived/synthetic_ohca_hotspots.csv. The file is a proxy dataset that stands in for real NSW Ambulance incident data until that feed lands in the project; every number downstream of this chapter — the Chapter 3 spatial maps, the Sprint 20 choropleth and proportional-symbol figures, the base-placement shortlist inputs — descends from the formula set out here. The output is therefore explicitly pending verification against NSW Ambulance data access; this status is propagated through the CSV’s data_label column (value synthetic) and through the run log.

Motivation

An earlier version of the synthetic cohort (Sprint 01 checkpoint, later retained through Sprints 16–20) ranked SA2s by an age-dominated composite proxy. That proxy ordered three wealthy northern-shore SA2s — Chatswood (East)–Artarmon, Hurstville, and Wahroonga–Warrawee — at the top of the list. Clinically, that ordering is implausible: OHCA incidence is driven by age and socio-economic disadvantage, not by age alone. Wealthy suburbs carry older age structures but markedly better cardiovascular health profiles (higher preventive-care uptake, lower smoking prevalence, better access to primary care) than the disadvantaged south-west corridor — Fairfield, Cabramatta, Liverpool, Bankstown, Mount Druitt, Campbelltown — where absolute OHCA counts should be higher. Sprint 26 replaces the old composite with an explicit SEIFA-weighted formula so the proxy reflects this.

Formula

For each SA2 i:

OHCA_i ∝ population_i × age65+ share_i × IRSD weight_i

where the IRSD weight is a monotonically decreasing function of the ABS 2021 SEIFA Index of Relative Socio-economic Disadvantage (IRSD) decile [37]:

$$\text{IRSD weight}_i = \frac{11 - \text{IRSD decile}_i}{10}$$

Decile 1 (most disadvantaged) maps to a weight of 1.0; decile 10 (least disadvantaged) maps to 0.1. A linear decile weighting is chosen for v1 because it is interpretable, monotone, and does not require committing to a specific published effect-size estimate — the published OHCA-by-deprivation gradient literature reports relative risks in the 1.5×–3× range between most- and least-advantaged quintiles (Australian and international cohorts), and a linear decile weighting sits conservatively inside that envelope. A log-linear or quintile-step weighting can be substituted when the NSW Ambulance cohort is joined in and the empirical gradient can be measured directly.

The product population × age65+ share gives the expected count of at-risk person-years; the IRSD weight then modulates that by cardiovascular-risk profile. Formally the proxy is a multiplicative hazard on person-years, not an age-standardised rate — the latter requires the NSW Ambulance event-level data.

Scaling and base-rate anchor

The per-SA2 raw weight is turned into an integer event count by linear scaling:

$$\text{scale} = \frac{T}{\sum_i \text{pop}_i \times \text{age65+}_i \times \text{IRSD}_i} \qquad \text{OHCA}_i = \text{round}\left(\text{raw}_i \times \text{scale}\right)$$

The target total T = 4 200 cases per year is chosen so the sum across the 37 project SA2s lands inside a [3 500, 5 000] envelope consistent with:

• Hasselqvist-Ax et al., 2015 [6] — the Swedish national registry paper reports a crude OHCA incidence of approximately 100 cases per 100 000 person-years, the closest well-characterised incidence we trust without a bespoke Australian fit.
• NSW Ambulance Annual Report 2022–23 [25] — reports roughly 4 800 adult-attended OHCA events per year across the Greater Sydney footprint. The project 37-SA2 subset covers a population substantially smaller than the full metro, so a 4 200 target is the upper-mid point of the envelope after the ABS 2021 Greater Sydney denominator [27] is allowed for.

Tessa’s Sprint 26 test test_csv_total_count_is_within_nsw_ambulance_envelope fails the build if the emitted total falls outside the [3 500, 5 000] window — the invariant is enforced, not just documented.

Inputs and provenance

All inputs live in data/curated/sa2_greater_sydney.csv and are labelled observed+estimated in that file’s data_label column. The “estimated” component reflects that the SEIFA decile values carried in the curated CSV are the project’s best reconciliation of the ABS IRSD SA2 datapack pending a full re-fetch; values flagged pending-verification in the curated table would cascade into the synthetic CSV as zero-weighted rows rather than hand-filled numbers.

Top-5 SA2s after re-ranking

The re-ranked top of the table shifts decisively into Sydney’s south-west — the pattern Roger anticipated in the Sprint 26 brief when he observed that “there are lots of wealthy boomers in the northern suburbs, but unhealthy people in the west / south west.”

All five SA2s sit in IRSD decile 1 or 2 (most disadvantaged). Wahroonga–Warrawee (IRSD decile 9) falls from rank 3 to rank 33, Chatswood (East)–Artarmon (decile 8) falls from rank 1 to outside the top-10, and the absolute counts redistribute mass toward the south-west — visible in the Chapter 3 proportional-symbol map once it is regenerated from this CSV.

Limitations and pending verification

1. Ecological fallacy. A monotone IRSD weight captures the average deprivation gradient across an SA2 but cannot resolve within-SA2 variation. Empirical Bayes smoothing at SA1 granularity is the right refinement, and is deferred to the sprint that consumes the NSW Ambulance incident data.
2. Age mid-point only. The proxy uses the 65+ share as a single breakpoint. The real OHCA age curve is steeper in the 75+ band; once the NSW Ambulance data is joined, per-year age bins should replace the 65+ share.
3. No sex stratification. Male OHCA incidence is roughly 1.7× female in the comparable age bands [6], but the ABS SA2 sex split is close to 50/50 in the 37 project SA2s so the aggregate ranking is insensitive to this. The model should still be re-fit once a stratified NSW Ambulance series is available.
4. Linear decile weighting is a placeholder. The v1 linear weighting is deliberately conservative and has no empirical fit. When the NSW Ambulance cohort lands, re-fit f(IRSD) directly on the joined dataset.
5. Output remains a proxy. Every downstream figure, table, and shortlist entry must carry the “synthetic cohort, pending NSW Ambulance data access” caveat until the real feed replaces this CSV. The Chapter 3 spatial-analysis caveat paragraph points readers back to this chapter.

Reproduction

The CSV is regenerable from the curated inputs with a single-file Python routine:

count_i = round(
    population_i * (age_65_plus_pct_i / 100) *
    ((11 - irsd_decile_i) / 10) * scale
)

where scale is calibrated so the integer totals sum to the target T = 4 200. Tessa’s Sprint 26 test test_counts_are_consistent_with_formula inverts this — it re-derives the scale from the dataset’s own totals and fails the build if any row deviates from formula prediction by more than one count.

Image Credits

This appendix records every image that appears in LIFT Study 01, its licence, and its source. It is generated from content/assets/images/manifest/image-manifest.json by scripts/build-image-credits.js; manual edits here will be overwritten on the next build. Corrections flow through the manifest.

Every image also carries alt text in the manifest; that alt text is wired into the published HTML and paged.js PDF via Paige’s templates so that screen-reader users receive a specific, clinical-register description of each figure’s content.