Bayesian Causal Atlas · Structural Causal Model
A Judea Pearl structural-causal decomposition of wasting-syndrome interventions — sourced hazard ratios and trial effect sizes, backdoor/front-door-structured mediation, and a desired-risk-reduction Pareto engine for the minimum effective set.
Cachexia is a multi-organ wasting syndrome driven by the underlying disease (advanced cancer, heart failure, COPD, CKD). This tool is a research-grade evidence aggregator, not medical advice and not a treatment protocol. Critically: almost no cachexia-directed therapy has demonstrated a mortality benefit — most move only proximal endpoints (weight, lean mass, appetite). The disease burden is a confounder that drives both the wasting and death; reductions shown here are upper-bound ceilings on the proximal endpoint, not survival guarantees. Several agents (e.g. megestrol) carry real harms. Decisions belong with the treating oncology / cardiology / palliative team.
The structure below is why cachexia is hard. The disease burden node U is a backdoor confounder: it drives systemic inflammation (and thus the wasting cascade) and drives mortality directly. Cachexia drugs act on the front-door mediators (inflammation → intake/catabolism → mass → function). If U keeps progressing, the front-door path is blocked at the exit — which is exactly why anamorelin lifts lean mass but not survival.
Adjustment strategy. Backdoor path A ← U → Mortality is closed by conditioning on disease-directed control (node R, "treat the driver"). Front-door identification along A → inflammation → mass → function isolates each agent's mediated effect, removing the spurious correlation between "patients who get drug X" and "patients whose tumour happens to be indolent."
Set your target reduction in cachexia-progression risk. The engine performs a greedy Pareto search — adding interventions by marginal Pearl-adjusted contribution (redundancy-discounted within mechanism class) — and returns the smallest set that reaches the target. Apply it to the table, then tune by hand.
Each row is a node in the SCM. Toggle interventions and drag the dose slider (dose scales the effect toward its saturation ceiling). The combined readout above updates live — this is the what-if engine. RRR = relative reduction in cachexia-progression risk (proximal endpoint), an upper-bound ceiling derived from the cited trial effect size and discounted by evidence grade.
| Intervention | Class | Proximal effect (sourced) | RRR ceiling | Grade | Dose | Ref |
|---|
Mechanism classes: ROOT disease-directed · INFLAM anti-inflammatory / GDF-15 axis · OREX orexigenic / appetite · ANAB anabolic / muscle · SUBSTR nutritional substrate · NEURO neurohormonal. Effects within a class are correlated and redundancy-discounted; across classes they combine near-independently.
Greedy marginal contribution · target line tracks slider
—
Leave-one-out drop in combined RRR
Each bar is the counterfactual loss of combined risk reduction if that single intervention were removed from the current set — its necessity. Small bars indicate redundancy (another agent in the same class is covering it).
The two parameters below control how much shared causal pathway is removed when interventions overlap. ρ-within discounts agents that hit the same mechanism (e.g. two orexigenics fighting over the same appetite circuit); ρ-across applies a residual haircut between classes for the shared inflammatory backbone.
how much overlap to subtract
What % of cross-correlation is appropriate? For cachexia, the inflammatory backbone (GDF-15 / IL-6 / TNF-α) is shared by most agents, so a defensible ρ-within sits at 0.40–0.70 (we anchor 0.55). ρ-across is smaller (0.10–0.25) because nutritional substrate, anabolic stimulus and disease control act through genuinely distinct mechanisms. Setting both to 0 (full independence) over-counts and is the optimistic extreme; the tornado at right shows the swing.
current selection across correlation extremes
—
propagating each trial's confidence interval
—
point estimate & dispersion
Each intervention's effect is drawn from a truncated normal whose spread is set by its evidence grade (A σ≈0.20·r, B 0.28, C 0.38, D 0.50 — wider for weaker evidence). The draws are recombined through the same noisy-OR each iteration, so the interval reflects both trial imprecision and structural correlation, not a naïve sum.
Cachexia agents saturate — and some reverse above their optimum. The dose sliders in §03 follow these curves: effect climbs toward a ceiling, then plateaus.
modelled effect vs dose fraction
sourced dose findings
| Ponsegromab | Monotonic 100→200→400 mg: +1.22 / +1.92 / +2.81 kg vs placebo — still ascending at 400 mg, no plateau reached in the GDF-15-high population [10]. |
|---|---|
| Omega-3 (EPA) | Inverted: 2 g/day showed a clinically relevant signal; 4 g/day gave no added benefit — saturation / tolerability ceiling [8]. |
| Olanzapine | Low-dose 2.5–5 mg is optimal; higher antipsychotic doses add sedation/metabolic toxicity without appetite gain [2]. |
| Megestrol | ≈800 mg/day optimal; higher doses trended to weight loss and raised thromboembolism — a true reversal [9]. |
| Anamorelin | 100 mg plateaus lean-mass gain (~1 kg); handgrip never responds at any dose — a ceiling on the functional axis [3]. |
Weight and lean mass are surrogates. Anamorelin (Grade A for lean mass) moved neither handgrip nor survival [3]; MENAC stabilised weight but not muscle or activity [6]. A "30% RRR on cachexia progression" may buy no survival or function. The honest headline metric for patients is QoL and independence, which most agents barely touch.
Front-door identification assumes the mediator (inflammation→mass) fully transmits the effect and that U doesn't have an unblocked direct path to the endpoint. In refractory cachexia the tumour drives wasting through pathways no current drug touches, so the adjustment is optimistic — the real ceiling is lower than 85%.
Breakthrough hook: the only consistent mortality mover is controlling the driver — effective anticancer therapy or guideline-directed HF therapy (ACEi/β-blocker reduce weight-loss risk and death [11,12]). The ROOT node dominates the Pareto frontier; everything else is supportive. A tool that ranks supportive agents must not obscure that hierarchy.
| Endpoint | Composite cachexia-progression risk: probability of continued clinically significant weight / lean-body-mass loss + anorexia persistence over a 12–16 week horizon. Not overall-survival HR (which is mostly null for these agents). |
|---|---|
| RRR ceiling derivation | For each agent, the published proximal effect size (Δ weight kg, Δ LBM kg, responder %, or weight-loss-risk reduction) is normalised against the placebo-arm natural history and capped, then multiplied by an evidence-grade factor (A 1.0, B 0.85, C 0.70, D 0.45). Result framed as an upper bound, not an expected value. |
| Dose model | Effect(dose) = ceiling × saturating function with agent-specific shape; EPA and megestrol use a non-monotone (peaked) curve to encode reversal at high dose. |
| Backdoor adjustment | Disease-burden confounder closed by conditioning on the ROOT node (disease-directed control). The ROOT node carries the only effect with a credible survival path. |
| Front-door combination | Within class: rank by effect, k-th agent contributes r·(1−ρ_within)^(k−1). Class aggregate R = 1−∏(1−r_eff). Across classes: rank by R, j-th class contributes R·(1−ρ_across)^(j−1); combined = 1−∏(1−R_eff). Capped at 0.85. |
| Pareto search | Greedy: repeatedly add the unselected intervention with the largest marginal increase in combined RRR; stop at target or when no positive marginal remains. Ties broken by lower patient burden. |
| Monte-Carlo | 4,000 iterations; each r_i ~ TruncNormal(r_i, (grade σ)·r_i) on [0, ceiling]; recombined via the noisy-OR each draw; report mean, SD, 2.5/97.5 percentiles. |
| Cross-correlation default | ρ_within = 0.55, ρ_across = 0.15 (defensible range 0.40–0.70 / 0.10–0.25). Independence (0/0) is the optimistic extreme. |
| Limitations | Effect sizes pool heterogeneous tumour/HF populations and definitions (Fearon 2011 vs ICC vs weight-loss-only). Surrogate endpoints. No head-to-head trials for most pairings — combination effects are modelled, not measured. |