VAST · Terra · C5 — citations

Vastopolis outbreak investigation. Method-level enrichment table (Stage 2 of the C3 chat; structure held fixed).

← Back
C5 Method-level enrichment table (Stage 2 of the C3 chat; structure held fixed)

Source-paper abbreviations

Method enrichment for the fixed workflow

Phase or step Analytical purpose Specific implementation methods and choices Source paper(s) Parameters and validation considerations
1. Extract a credible flu-like-illness signal from posts Separate likely contemporaneous illness reports from news, jokes, historical references, and other irrelevant posts. Implement a text-classification pipeline that assigns each post an illness relevance score, symptom indicators, report type, and exclusion reason. Present the resulting posts in linked map, timeline, and text views, allowing analysts to inspect examples and revise vocabulary, thresholds, and exclusion rules. The source papers do not specify NLP or syndrome-classification methods. The linked, interactive inspection pattern is supported generally by [ST-Model] and [ISeeCube]. Create a manually reviewed validation sample stratified by illness score, time, and zone. Record precision/false-positive patterns separately for first-person reports, third-person reports, news, and figurative language. The classifier threshold should be adjusted based on this review, not selected solely by model accuracy.
1. Signal-quality refinement loop Ensure apparent growth is not primarily due to posting volume, a viral news story, or an extraction artefact. Use coordinated filtering and highlighting: selecting a time interval or zone in one view updates all others. Show score distributions, representative messages, and the spatial distribution of retained versus excluded posts. Maintain an audit trail of each revised signal specification. [ST-Model] recommends coordinated map/time-graph views, spatial/temporal/attribute filters, and immediate visual feedback during refinement. [ISeeCube] demonstrates a coordinated overview-plus-detail design: a space-time overview identifies periods of interest, while timelines and detailed views support interpretation. Compare results before and after each classifier revision. Check whether the leading candidate origin areas remain stable under reasonable changes in the inclusion threshold. If they do not, report the instability rather than treating the signal as a robust outbreak indicator.
2. Regularise and aggregate posts into zone-time units Produce population-adjusted spatial time series suitable for outbreak detection and mechanism comparison. Aggregate high-confidence illness posts by the 13 supplied zones and a chosen time interval. Compute: raw post count, confidence-weighted count, rate per resident population, onset time, growth rate, persistence, and symptom composition. Transform absolute counts to relative values before comparing zones of unequal population. [ST-Model] explicitly recommends transforming absolute values into relative values, such as dividing by population, before analysis. [Places] recommends aggregating events by place and time interval and calculating counts and attribute statistics for each place-time combination. Choose a time-bin duration based on the reporting frequency and response needs. The papers use one-hour bins in mobility examples, but do not prescribe a health-surveillance bin. Test, at minimum, a fine and coarse resolution—for example, 6/12-hour and daily bins—and check whether onset ordering changes. Use both raw counts and population-normalised rates; small-population zones can otherwise appear artificially extreme.
2. Define candidate outbreak locations / hotspots Distinguish sustained, spatially coherent illness concentrations from isolated posts. For the existing zone-based workflow, first identify zones with sustained excess burden. If post-level hotspot delineation is needed, use density-based clustering rather than k-means: DBSCAN or OPTICS can identify irregularly shaped dense concentrations and leave isolated points as noise. A two-stage procedure can first cluster in space-time, then spatially cluster retained events to combine repeated episodes at the same location. [Places] Density-based clustering needs: spatial threshold, temporal threshold, optional thematic thresholds, and minimum density. [Places] suggests estimating a spatial threshold by measuring distances to the third or fourth nearest neighbours in visible concentrations. Refine thresholds iteratively in changes of roughly 10–25% when clusters are implausibly fragmented or over-merged. Its illustrative traffic setting used 100 m, 10 min, and 20° direction thresholds; these values are not transferable to Vastopolis.
2. Choose cluster distance definitions where post-level clustering is used Avoid joining reports that are geographically close but temporally unrelated, or that represent different relevant conditions. Use a custom multidimensional distance that always includes space and may include time, symptom class, report confidence, or other meaningful features. For cyclic attributes, such as wind direction or hour of day, use cyclic distance rather than direct numeric subtraction. Prefer a maximum-based combination of normalised component distances when the analyst wants every threshold constraint to be satisfied. [Places] In [Places], a pair is excluded if any component distance exceeds its threshold. The paper prefers the maximum-based formulation over Euclidean combination because the latter can reject pairs that are individually within all thresholds but jointly near their limits. Select thresholds from domain assumptions, data resolution, and interactive inspection.
3. Derive contextual features Characterise whether observed propagation is associated with adjacency, population, highways, water bodies, hospitals, landmarks, or wind. Build zone-time features for: zone adjacency; distance or intersection with highways and water bodies; proximity to hospitals and landmarks; population; wind speed/direction; and wind-aligned relationships between possible source and recipient zones. Preserve the raw spatial-context layers so analysts can inspect whether an apparent relationship is driven by one feature or a few unusual zones. [Places] supports deriving dynamic attributes representing relationships to spatial, temporal, and spatio-temporal context, including distances, neighbourhood counts, and relations within spatial or temporal windows. The supplied data allow proximity and directional-association measures, but do not contain water-network flow, environmental pathogen measurements, indoor ventilation, confirmed exposure locations, or contact data. Therefore, proximity to water or wind alignment is only hypothesis-supporting evidence, not direct proof of waterborne or airborne transmission.
3. Explore burden, onset, and context visually Let analysts recognise early clusters, onset order, possible spread fronts, and reporting artefacts. Use linked views comprising: (1) a choropleth/proportional-symbol map; (2) zone time-series plots; (3) an onset-ranking view; (4) a space-time representation or animated/small-multiple map; and (5) selectable contextual layers for highways, water, hospitals, and wind vectors. Provide brushing, time-range filtering, zone selection, and simultaneous highlighting across all views. [ST-Model]; [ISeeCube] [ST-Model] uses map displays linked to time graphs, with spatial, temporal, attribute, and cluster filters. [ISeeCube] supports using a space-time overview to find non-sequential periods of interest, then inspecting them in linked timeline/detail views. The specific gaze-density and scanpath methods in [ISeeCube] are not applicable to this task.
3. Summarise numerous zone time series Prevent overplotting while retaining variation among zones. Show individual zone series when there are few; otherwise show distribution summaries per time step using median/mean plus quantile bands or a temporal histogram. Cluster zones with similar temporal burden profiles, then map cluster membership and examine whether the resulting spatial pattern is epidemiologically interpretable. [ST-Model] [ST-Model] uses mean lines and quantile summaries—quartiles, quintiles, or deciles—to show multiple series without clutter. It tests several values of k for k-means, reviews map and time-graph results, then progressively subdivides internally heterogeneous clusters. There is no universal correct k; select it by within-cluster temporal homogeneity and interpretable spatial distribution.
4. Calibrate lagged mechanism comparisons Examine whether changes in a plausible source/exposure series precede changes in a recipient illness-rate series at meaningful lags. For each candidate mechanism, define an interpretable bivariate pair: for example, a source-zone burden or exposure index as the base series and a possible recipient-zone burden as the response series. Compute impact scores across sliding windows and a range of lags. Use the tolerance-based Kendall’s tau as a robust default; compare it with standard Kendall, Spearman, or Pearson correlation when useful. [Cross-Impact] The method assumes regularly sampled series with a shared temporal resolution. Resample or impute missing values before analysis. The paper recommends initially setting tolerance to about 3–5% of the series range, increasing it for noisier data and decreasing it for subtler patterns. Its COVID case study used 2% tolerance, a 7-day window, and 7–21-day lags; these are case-specific starting points, not flu defaults.
4. Select time windows, lag ranges, and significance thresholds Balance sensitivity to short-lived change against noise reduction and avoid searching implausible delays. Choose the sliding-window length to align with meaningful temporal cycles, particularly weekly reporting cycles. Set maximum lag from public-health knowledge of plausible symptom, reporting, intervention, and environmental-response delays. Use visual inspection and sensitivity analysis to compare alternative windows and lag ranges. [Cross-Impact] Larger windows smooth fluctuations but can obscure short events; shorter windows reveal transient effects but are noisier. For standard Kendall’s tau without ties, the paper provides approximate two-tailed p = 0.05 critical values: (
4. Inspect lagged patterns interactively Determine whether an apparent mechanism is stable, episodic, lag-specific, or noise. Implement an Ikat-style plot: horizontal axis = absolute time, vertical axis = lag, colour = positive/negative lagged association. Link a hovered cell to the corresponding base and response time-series segments and a scatterplot. Add dynamic queries for time range, level, trend, lag, and impact strength. [Cross-Impact] Interpret vertical bands as sustained effects across a range of lags; diagonal structures as transient effects whose apparent lag changes through time; horizontal bands as fixed-lag relationships; isolated patches as episodic effects. Treat these as exploratory patterns. The source paper explicitly frames them as temporal associations, not causal proof.
4. Extract mechanism-support events Convert visually identified stable lagged associations into inspectable, comparable objects. After analysts settle settings on representative zone pairs, automatically extract contiguous regions in the thresholded time-lag matrix. Store each event with base interval ([t_1,t_2]), lag interval ([l_1,l_2]), duration, lag depth, impact strength, and base/response trend descriptors. [Cross-Impact] The source implements event extraction by finding four-neighbour connected regions in a Boolean matrix of cells satisfying all selected conditions. Enforce minimum duration and minimum lag depth. Its case study used a trend threshold of ±3% and (
4. Conduct parameter-sensitivity analysis Establish whether origin or mechanism findings are robust to reasonable analytical choices. Vary one parameter at a time around the current value—for example, tolerance, impact threshold, trend threshold, time-window size, or maximum lag. For each setting, recompute the number of extracted events, total event size, coverage of base/response series, leading source zones, and mechanism ranking. Inspect plots for bend points and unstable regions. [Cross-Impact] The paper defines a parameter range ([v-\eta, v+\eta]) and increment (\delta), then evaluates extraction at each setting. Bend points—where event count or coverage changes rapidly—identify values requiring closer visual examination. A conclusion that disappears under a small parameter change should be labelled low confidence.
5. Fit and compare origin and transmission hypotheses Estimate which origin locations and candidate mechanisms best explain observed propagation. Use separate, interpretable mechanism-specific models rather than a single opaque global model. Candidate predictors can include prior burden in adjacent zones, wind-aligned exposure, water proximity, or common-source proximity. Group zones with similar temporal profiles where appropriate, build models for representative series or groups, and retain a clear description of each model’s assumptions and parameters. [ST-Model] supports clustering spatial time series before modelling, modelling representative series, and retaining explicit model descriptions. [Cross-Impact] supports interpretable lagged association and event features. The uploaded papers do not provide a validated person-to-person, airborne, or waterborne disease-transmission model. In particular, the papers do not justify causal route attribution from microblogs, static water features, or wind observations alone. Any route-specific model must therefore be specified and reviewed by epidemiologists, and its outputs presented as ranked hypotheses.
5. Build representative time series and choose model forms Avoid fitting highly unstable models to every individual zone series while preserving meaningful heterogeneity. For a group of similar zones, construct a representative series from the mean, median, or another percentile at each time step. Optionally trim a selected percentage of highest and lowest values before computing means. For periodic variation, consider Holt-Winters/triple exponential smoothing; for relationships between an exposure index and burden, use interpretable linear or polynomial regression as exploratory baselines. [ST-Model] [ST-Model] advises comparing the representative curve to individual series. Holt-Winters requires an explicit cycle length and at least two full cycles for model construction. Linear/polynomial regression is appropriate only when the chosen exposure-response form is substantively plausible. Do not assume periodicity from a few days of outbreak data.
5. Evaluate model fit and residuals Identify systematic mismatches that indicate a wrong mechanism, missing covariate, bad grouping, or reporting bias. Visualise observed versus predicted burden, residual maps, residual time graphs, and residual distributions. Examine whether residuals show temporal regularity, spatial concentration, or consistent sign in certain zones. If they do, revise the model or subdivide the relevant zone group. [ST-Model] The key criterion is not merely small residual magnitude: residuals should lack clear systematic spatial or temporal patterns. [ST-Model] recommends residual time-series summaries using quantiles and temporal histograms, then clustering residual series to identify subgroups that need separate models. Store residual mean, standard deviation, and distribution statistics with the model.
5. Avoid over-complex global spatial models Prevent arbitrary spatial-weight assumptions from creating a misleading “best” transmission explanation. Prefer a small set of transparent candidate models and compare their predictions and residual patterns. If using a global spatial-temporal model such as STARIMA, document and justify every spatial-weight matrix, lag, and neighbourhood rule. [ST-Model] [ST-Model] notes that global models such as STARIMA depend critically on analyst-specified spatial weights and can behave as difficult-to-interpret black boxes. The paper reports literature where independently fitted local temporal models outperformed a global STARIMA model when spatial weights were simplistically specified.
6. Create an operational evidence summary Translate results into response priorities while preserving uncertainty and alternative explanations. Produce a ranked evidence package containing: candidate origin areas and onset times; mechanism-specific evidence; sensitivity results; residual anomalies; zones needing field investigation; and data needed to discriminate between hypotheses. Use event timelines to show when and where candidate impacts occurred, and map selected event sets back to geography. [Cross-Impact] supports timeline-based analysis of extracted lagged-impact events. [ST-Model] supports externally storing model descriptions, parameters, group membership, and residual statistics. Preserve reproducibility: store the illness-detection specification, aggregation interval, population denominator, clustering settings, lag/window/tolerance settings, model parameters, and analyst interpretations. Recommendations should be conditional: for example, “prioritise water sampling if water-associated patterns persist after controlling for population and reporting activity.”
6. Confirm or reject hypotheses through external investigation Prevent social-media-derived patterns from being mistaken for confirmed epidemiological causality. Use the visual-analysis outputs to direct clinical case investigation, laboratory confirmation, environmental testing, and contact/exposure inquiry. Compare inferred origin and route patterns against those independent sources before deploying high-consequence interventions. No supplied source paper provides public-health confirmation protocols. The limitation follows from the scope of all four papers: they support exploratory visual analytics, clustering, modelling, and lagged-pattern analysis rather than causal proof. Treat microblog evidence as a surveillance proxy. A waterborne claim requires relevant environmental and epidemiological evidence; an airborne claim requires appropriate exposure/environmental evidence; and a person-to-person claim requires case/contact evidence.

Practical parameter starting plan

These are initial exploratory settings, not fixed defaults:

Component Initial settings to test Rationale
Zone-time aggregation 6-hour, 12-hour, and daily bins Tests whether onset ordering is robust to temporal resolution.
Smoothing Unsmoothened and a short moving average; avoid relying only on a 7-day average during an acute event [Cross-Impact] used a 7-day moving average for COVID reporting effects, but this may hide a rapidly emerging flu outbreak.
Lagged cross-impact window 7 and 14 daily observations, if enough daily history exists [Cross-Impact] recommends windows aligned with natural cycles; weekly cycles are common in reporting and human activity.
Lag range A domain-reviewed range covering plausible reporting and transmission delays [Cross-Impact] recommends setting maximum lag from domain knowledge, not indiscriminate search.
Kendall-tau tolerance Start at 3–5% of each series’ robust range; compare with 2% and 0% [Cross-Impact] recommends 3–5% as a starting range and used 2% in its COVID case study. Use a percentile-based range, such as 5th–95th percentiles, if outliers are influential.
Impact threshold Explore ( \tau
Cluster refinement Change spatial/temporal thresholds in 10–25% increments Direct recommendation from [Places] for judging under- versus over-merged density clusters.

Possible refinements to the workflow structure

These are not incorporated into the fixed workflow, but the source papers suggest they could be useful.

  1. Add an explicit data regularisation stage before lagged analysis.
    [Cross-Impact] assumes regularly sampled time series with common temporal resolution. A dedicated step could handle missing periods, resampling, reporting-cycle smoothing, and documentation of imputation before mechanism analysis.

  2. Add a standalone impact-event extraction and event-analysis phase.
    [Cross-Impact] converts stable lagged patterns into explicit impact events with duration, lag range, trend, and impact-strength attributes. Such events can then be compared across zones, time periods, and candidate mechanisms rather than remaining only as model outputs.

  3. Add parallel data-derived hotspot delineation alongside the fixed 13-zone analysis.
    [Places] argues that predefined territorial units can obscure irregular, overlapping, or elongated meaningful places. A parallel post-level DBSCAN/OPTICS analysis could identify candidate outbreak hotspots that cross zone boundaries, then compare them with the zone-based results.

  4. Add an explicit model archive and replay capability.
    [ST-Model] stores model type, parameter values, initialisation values, group membership, and residual statistics in an external human- and machine-readable representation. This would strengthen reproducibility during an evolving emergency response.

  5. Add a formal external-evidence integration stage.
    This is not supplied by the four visual-analytics papers, but it is necessary for public-health validity. Clinical reports, testing data, environmental samples, mobility/contact data, and water-system information would allow the current hypothesis-ranking workflow to become a confirmatory epidemiological investigation.