Temporal autocorrelation patterns in WHO GLASS country-level antibiogram time series encode early-warning signatures of horizontal gene transfer (HGT) events for specific resistance genes up to 18 months before phenotypic resistance becomes clinically detectable at threshold prevalence. An LSTM trained on per-country GLASS resistance frequency vectors will achieve >80% sensitivity for novel MDR emergence in held-out national datasets, because HGT events produce characteristic low-amplitude oscillations in multiple unrelated drugs simultaneously — a signature invisible to single-drug surveillance but detectable as a correlated cross-drug anomaly in multivariate time series analysis.

An LSTM (or LSTM-class recurrent) model trained on multivariate, per-country, per-pathogen-drug resistance-frequency time series from WHO GLASS (quarterly or annual resolution, ≥5 drugs per pathogen, ≥6 years history) can classify, in held-out countries and held-out time windows, whether a specific pathogen-drug combination will cross a clinically defined "emergence" threshold (e.g., resistance prevalence rising from <10% to ≥20% within a 24-month window) with sensitivity >80% and a median lead time ≥18 months before the threshold-crossing timestamp, using only pre-crossing multivariate correlated-oscillation features across nominally unrelated drugs as predictors, at a false positive rate (specificity-complement) no worse than 1 alarm per country-pathogen series per 5 years.

Minimum viable test (MVT), ~6-8 weeks, single analyst + 1 ML engineer:

1. Extract all GLASS country×pathogen×antibiotic annual resistance-frequency series 2016-2023 (public WHO GLASS data portal).
2. Define emergence-event label: prevalence crosses from <10% to ≥20% within any 24-month window, per pathogen-drug-country triple; require ≥3 prior years of data to qualify as a "predictable" event.
3. Hold out entire countries (not just time windows) for test set — geographic block cross-validation (leave-region-out), 5 folds by WHO region.
4. Train (a) proposed multivariate LSTM using all drugs/pathogens per country as joint input channels, and (b) two baselines: univariate ARIMA/exponential-smoothing per drug, and gradient-boosted trees on hand-engineered lag/slope/volatility features.
5. Score sensitivity, specificity, lead-time distribution, and AUROC/AUPRC at matched operating points across all three models.
6. Run ablation: shuffle drug-channel identity (destroy true cross-drug correlation structure but preserve marginal statistics) — if performance is unchanged, the "HGT cross-drug signature" claim is falsified regardless of raw accuracy.
7. Where WGS/plasmid surveillance data exists for a subset (e.g., ECDC EARS-Net linked genomic studies, or published outbreak reports), manually cross-check whether flagged "early warning" windows coincide with documented HGT events versus clonal spread.

# 1. Data prep
tensor[country][pathogen][antibiotic][year] = resistance_frequency  # from GLASS
labels[country][pathogen][antibiotic] = emergence_event(t0, threshold=0.20, from_below=0.10, window=24mo)

# 2. Baselines
for each (country,pathogen,antibiotic):
    fit_ARIMA(series)              -> baseline_1_score
    fit_GBT(lag_features(series))  -> baseline_2_score

# 3. LSTM model
class MDRLSTM(nn.Module):
    def __init__(self, n_drugs, hidden=32, layers=2):
        self.lstm = nn.LSTM(input_size=n_drugs, hidden_size=hidden,
                             num_layers=layers, dropout=0.2, batch_first=True)
        self.head = nn.Linear(hidden, 1)  # sigmoid -> P(emergence within 24mo)

    def forward(self, x):              # x: [batch, lookback_periods, n_drugs]
        out, _ = self.lstm(x)
        return torch.sigmoid(self.head(out[:, -1, :]))

# 4. Cross-validation
for region in WHO_REGIONS:
    train = countries not in region
    test  = countries in region
    model = train_LSTM(train, class_weight=inverse_freq(labels))
    eval_sensitivity_specificity_leadtime(model, test)

# 5. Critical ablation
shuffled_x = shuffle_drug_channels(x)   # break cross-drug correlation, keep marginals
ablation_score = eval(model_trained_on_shuffled, test)
assert true_channel_score - ablation_score >= 10pp   # else core claim falsified

# 6. Mechanistic check
for flagged_window in early_warnings:
    check_overlap(flagged_window, genomic_HGT_events_from_NCBI_PathogenDetection)