Continual meta-learning of personalized neural surrogates via dynamic context management architectures reduces context ballooning and compute costs in personalized cardiac electrophysiology simulations.

A continual meta-learning system that uses dynamic context management (adaptive selection, compression, and eviction of patient-specific context tokens/embeddings across sequential personalization episodes) will, for personalized cardiac electrophysiology (EP) neural surrogate models, achieve: (a) ≥90% of the predictive accuracy (measured as RMSE on transmembrane potential / activation time maps vs. ground-truth monodomain/bidomain PDE solves) of a full-context (non-pruned, cumulative) meta-learning baseline, while (b) reducing peak context size (tokens/embeddings retained in memory) by ≥50% and (c) reducing wall-clock/GPU-hour compute cost of per-patient fine-tuning or inference by ≥40%, measured across a sequence of at least 20 distinct simulated or real patient adaptation episodes with no catastrophic forgetting (performance degradation on earlier patients <10% after 20 sequential updates). The hypothesis is falsified if any one of (a)-(c) fails to hold simultaneously under matched compute/accuracy budgets.

Minimum viable test (MVT): Use a synthetic cohort of N=30 virtual patients generated via parametric perturbation of a reference bi-ventricular EP model (varying conductivity, fiber angle, scar location/size). Train (1) a full-context cumulative meta-learner, (2) the proposed dynamic context-managed meta-learner, and (3) two heuristic baselines (fixed-window, random eviction) sequentially over the 30-patient stream. Evaluate accuracy, context size, compute cost, and forgetting after each of the 30 adaptation steps. Repeat with 3 random task orderings and 3 seeds (9 runs per method) to assess variance and order sensitivity.

# Pseudocode: Dynamic Context-Managed Continual Meta-Learning for Cardiac EP Surrogates

base_surrogate = pretrain_graph_neural_operator(generic_cardiac_meshes)
context_store = ContextManager(budget=B, method="learned_gating")

for patient_t in patient_stream:
    obs_t = get_sparse_observations(patient_t)  # ECG, limited EP measurements

    # Retrieve relevant compressed context from prior patients
    retrieved_context = context_store.retrieve(query=encode(obs_t))

    # Meta-learning adaptation step (e.g., MAML inner loop)
    theta_t = base_surrogate.clone_params()
    for step in range(K_inner_steps):
        loss = EP_loss(theta_t, obs_t, retrieved_context)
        theta_t = theta_t - lr_inner * grad(loss, theta_t)

    personalized_surrogate[patient_t] = theta_t

    # Evaluate against ground-truth PDE solve (held-out validation nodes)
    rmse_t = evaluate(theta_t, ground_truth_simulation(patient_t))
    log_metrics(rmse_t, context_store.size(), compute_time)

    # Dynamic context update: score, compress, evict
    new_embedding = encode(patient_t, theta_t)
    scores = context_store.gating_network(new_embedding, context_store.all_items())
    context_store.update(new_embedding, scores, budget=B)
    # compression: low-rank projection or clustering of retained embeddings
    context_store.compress_if_over_budget()

    # Periodically check forgetting on probe set
    if t % 5 == 0:
        forgetting_metric = evaluate_on_probe_set(personalized_surrogate, probe_patients)

# Outer meta-update (periodic, across batch of recent patients)
meta_update(base_surrogate, recent_patient_batch, context_store)

Key architectural components: (1) gating network (small MLP/attention) scoring context items by relevance+recency+information-gain, (2) compression via learned low-rank autoencoder or clustering (k-means on embeddings), (3) eviction policy enforcing hard budget B, (4) periodic outer-loop meta-update to adapt gating network itself based on downstream accuracy/forgetting signal (making context management itself meta-learned, not hand-tuned).