Wave-Guided Heat and the Diagnosis Tree: A Self-Controlled Study of Which Illnesses Follow Rossby-Wave Weather

Heatwaves and pollution-trapping high-pressure systems do not form in isolation. They often arrive along preferred routes steered by planetary-scale atmospheric waves (Rossby waves), whose energy can be tracked days in advance. Health studies of weather ignore this. They regress hospital admissions on the temperature measured locally, one diagnosis at a time, and they rarely check whether their confidence intervals can be trusted. We propose a single statistical model that addresses three of the resulting weaknesses. Each patient serves as their own control, which removes fixed between-person differences. The delay between weather and illness is learned rather than assumed. Rare diagnoses borrow statistical strength from related diagnoses through the existing medical code hierarchy, the tree that groups specific codes into broader categories.

The concrete question is which branches of the diagnosis tree show raised risk after wave-guided heat and stagnation episodes, at what delays, and whether the answer survives the empirical calibration that has exposed large biases in observational health research. The answer matters for two reasons. The atmospheric signal gives lead time for hospitals. The diagnosis tree is where rare but serious effects, such as specific arrhythmias or electrolyte disorders, currently disappear into noise.

Core Claim

The self-controlled case series (SCCS) is a design that uses only people who had the outcome and compares their event rate in exposed versus unexposed periods. We claim that SCCS, combined with a convolution over lags and a tree-structured prior over diagnosis codes, recovers weather–illness associations for rare diagnoses that flat, per-code analyses miss. We also claim it does so with calibrated confidence intervals.

The weather exposure is a set of pathway-specific indices built from Takaya–Nakamura wave activity flux (a vector field whose arrows show where wave energy is travelling). We compare these indices against local temperature and against empirical orthogonal function (EOF) modes, the principal components of gridded weather fields. The goal of that comparison is to test, not assume, whether the upstream dynamics add anything.

The novelty is not "put climate into epidemiology." It lies in two places. First, the explicit ICD hierarchy enters the SCCS likelihood as a shrinkage structure. Second, this structure is paired with learned lag profiles for a shared, exogenous environmental exposure and evaluated with negative and synthetic positive controls.

Where This Sits

Three prior results motivate the proposal. First, wave dynamics organize surface extremes. Wang et al. show that Northern Hemisphere heatwaves travel along four preferred pathways guided by enhanced wave activity flux: Asian, Western European, and two North American. Kou et al. trace flux from Atlantic blocking across Eurasia into the East Asian highs that trap pollution. In contrast, health studies such as the Chinese cardiovascular air-pollution study use only ground-level pollutant concentrations.

Second, rare codes are where signals are lost. Liu and Aickelin recover adverse drug reaction signals only after rolling Readcodes up to parent categories. The MIMIC coding replication finds that rare codes, not document length, are the bottleneck for every model it tests.

Third, observational estimates are badly miscalibrated unless checked. The OHDSI benchmark reports that only about half of nominal 95% intervals cover the truth, and that SCCS reaches the highest discrimination. The vaccine-surveillance evaluation finds that calibrated SCCS detects small effects fastest.

Two papers supply the method, and each lacks half of it. ConvSCCS learns piecewise-constant lag profiles using Group-Lasso and Total-Variation penalties, but it treats outcomes as flat and unrelated. Factorized SCCS shares strength across outcomes through latent factors learned from data, not from an ontology, and was tested on only 70 patients. We combine the lag convolution of the first with a fixed-ontology version of the sharing idea in the second.

One constraint shapes the whole design. SCCS controls only for characteristics that do not change over time. Weather is strongly seasonal, so season, influenza waves, holidays and day of week must be modeled explicitly. A convenient property works in our favor: illness cannot change the weather, so the SCCS assumption that events do not alter future exposure holds automatically.

Model

For patient $i$ with diagnosis $c$, the event rate on day $t$ is

$$\lambda_{ic}(t) = \exp\Big(\alpha_{ic} + s_{r(i)}(t) + \sum_{\ell=0}^{L} \beta_{c,\ell}\, x_{r(i)}(t-\ell)\Big),$$

The symbols are as follows. $\alpha_{ic}$ is a person-specific baseline, which cancels in the conditional likelihood. $s_{r}(t)$ is a calendar-time spline for the patient's region $r(i)$, together with day-of-week, holiday and influenza terms. $x_r(t)$ is the regional exposure index. $\beta_{c,\ell}$ is the log relative risk at lag $\ell$ days, with $L$ up to 21 days.

The hierarchy enters through the prior on the lag coefficients:

$$\beta_{c,\ell} = \beta_{\mathrm{pa}(c),\ell} + \delta_{c,\ell}.$$

Here $\mathrm{pa}(c)$ is the parent node in ICD-10, running leaf → three-character code → block → chapter. The deviation $\delta_{c,\ell}$ is penalized by a group penalty per code and a Total-Variation penalty across lags. The penalty strength is allowed to depend on tree depth, which addresses the uneven depth of different ICD branches.

The effect of this structure is simple. A leaf code with few events inherits its parent's lag profile unless its own data argue otherwise. A code with many events can depart from its parent.

Research Program

Exposures. From ERA5 reanalysis for 1979 onward, we compute daily horizontal Takaya–Nakamura flux at 300 hPa. From it we build, for each health region, a "wave-guided index": flux convergence over the region, conditioned on upstream flux along the Western European or North American pathways identified by Wang et al..

We build two competing exposure representations. The first is the leading Varimax-rotated EOFs of 500 hPa height anomalies, following Runge et al.. The second is extended EOFs that capture lagged evolution, as in Yu et al.. ERA5 local 2 m temperature serves as the local reference.

Health data. The primary setting is SNIIRAM hospital diagnoses in France, linked by département; France lies on the Western European pathway. The replication setting is the US CCAE and MDCR claims used in the OHDSI benchmark, linked by state, which lie on the North American pathways. Access to the US data is the main feasibility risk. If it fails, UK primary care data from THIN serve as a fallback, though coding there is coarser.

Outcomes. All ICD-10 hospital diagnoses whose chapter has at least 500 events enter the tree. Leaves are retained regardless of their event count.

Calibration. Negative controls come in two kinds. Negative control outcomes are diagnoses with no plausible weather link, such as incident cancer diagnoses and chronic congenital conditions. Negative control exposures are the future wave index at leads of 7 to 21 days, which cannot cause today's admissions. Synthetic positive controls are injected into real negative-control outcomes at leaf level, with relative risks of 1.5, 2 and 4 and known lag shapes, following the OHDSI protocol. All intervals are then empirically calibrated.

Baselines

Naive reference. The simplest reference is standard SCCS with local 2 m temperature above the regional 95th percentile as a binary exposure, a fixed 0–3 day risk window, and each leaf code fitted separately. This is what most weather–health analyses amount to, and it sets the floor. The OHDSI benchmark reports that SCCS adjusted for other drugs reaches AUC 0.95 to 0.98 on drug negative and positive controls (Schuemie et al.). That figure comes from drug exposures, not weather, so we treat it as the level a well-behaved design should approach rather than a number to transfer directly.

Environmental epidemiology standard. Regional time-series Poisson regression with a distributed-lag term on local temperature and the same seasonal splines is the most likely objection from an environmental epidemiologist. We include it because, for a shared exposure, SCCS and aggregate time-series models estimate similar quantities.

Flat ConvSCCS. ConvSCCS with the same lag penalties but no tree is the strongest existing lag method. No reported result for it on weather exposures or ICD-scale outcomes was found on Lacuna. Its paper reports only that it beats SmoothSCCS and NonParaSCCS in mean absolute error, without numbers in the file.

Latent-factor sharing. Factorized SCCS is the strongest method for sharing across outcomes. It reports AUC up to 85% for coefficient-sign recovery on AIM-94, against 69.3% for univariate SCCS.

Fixed roll-up. A roll-up to three-character codes followed by SCCS, in the style of Liu and Aickelin and Le et al., tests whether the tree prior is better than simply truncating codes.

Fairness and ablations. Every model uses the same cases, calendar adjustment, exposure series, lag range and a matched tuning budget, with penalties chosen by the same region-blocked cross-validation. Two ablations isolate the proposal's ingredients: the tree prior with fixed windows instead of convolution, and the tree prior with local-temperature exposure instead of the wave index.

What counts as winning. On leaf-level synthetic positive controls with RR 1.5, the tree model must show calibrated AUC and interval coverage better than both flat ConvSCCS and the fixed three-character roll-up. It must also show lower mean squared error on the injected lag profiles. On common codes, it must not do worse than flat ConvSCCS.

Why not just…?

Why not just use local temperature? This is the real threat. Local temperature is probably the proximate cause, and upstream flux may add nothing to effect estimation. The experiment settles the question directly. We fit the tree model with local temperature and with local temperature plus the wave index. If the wave index adds no fit beyond local temperature and heatwave duration, we report that openly. The contribution then narrows to the hierarchical lag method, with the wave index useful only for early warning.

We still test early warning as a separate claim. We compare how well ERA5-derived upstream flux at 3 to 7 days ahead predicts chapter-level excess admissions, against a persistence forecast of local temperature. We acknowledge that operational forecasts are strong: WeatherBench reports IFS 3-day Z500 RMSE of about 154 m²s⁻² versus 936 m²s⁻² for persistence (WeatherBench). A fair early-warning claim would eventually need archived forecasts, which the file does not supply.

Why not just use EOF modes as the exposure? EOFs are cheap and compress fields well, but orthogonality can mix separate physical processes into one mode. EOFs also cannot tell a propagating wave train from a standing block, and the flux diagnostic separates exactly these cases. EOF and extended-EOF exposures are run inside the identical model. If they match the wave index on held-out fit, we adopt them as the simpler exposure.

Why not just roll codes up one level, or let factors be learned? Truncation throws away real leaf-level differences, which is the specificity cost the file identifies. Learned factors need data to discover structure that the ontology already encodes, and Factorized SCCS was only tested on 70 patients. The fixed roll-up and Factorized SCCS baselines above measure both costs on the same injected leaf effects.

Why not just use a time-series or case-crossover design? For a regionally shared exposure, these designs estimate much the same thing, and we run the time-series version as a baseline. SCCS earns its place in two ways. It works with individual observation periods, handling enrolment and death censoring. It also lets age vary within a person. If the two designs agree after calibration, that agreement is itself a robustness result.

Decisive Checks

The first check is falsification. Future wave indices, used as negative control exposures, must give calibrated relative risks near 1. Any lead effect indicates residual seasonal confounding, and we refit with more flexible calendar terms before interpreting anything.

The second check is the leaf-recovery test on injected effects. It decides whether the tree prior is worth having.

The third check replicates in the US. Chapter-level lag profiles estimated in France should reappear in the North American databases for the corresponding pathway. They should appear there in shape, not necessarily in magnitude. A branch of the tree that shows raised risk in both settings and survives calibration counts as a finding. One that appears in only a single setting is reported as exploratory.