Wireless Sensor Networks (WSNs) deployed in real-world settings face heterogeneous operatingconditions in which conventional cluster-head (CH) selection based on instantaneousmetrics cannot anticipate short-term reliability degradation, leading to prematureCH failures and reduced network lifetime. We propose PSO-LSTM, a proactive reliabilitygatedextension of PSO-EECS, in which a lightweight Long Short-Term Memory (LSTM)model at the base station predicts short-horizon node failure probabilities from per-nodetemporal signals. These predictions form a set-level reliability score that multiplicativelygates the energy-centric term of the PSO fitness, recovering PSO-EECS behavior understable conditions. Experiments are conducted over 20 paired seeds on a 200-node heterogeneousWSN incorporating six co-occurring heterogeneity factors: energy variability, spatialhotspots, stochastic traffic, retransmission faults, energy harvesting, and asymmetric sinkplacement. Compared with seven bio-inspired baselines, namely PSO with GravitationalSearch Algorithm (PSOGSA), Grey Wolf Optimizer (GWO), Whale Optimization Algorithm(WOA), Ant Lion Optimizer (ALO), Harris Hawks Optimization (HHO), PSO-EECS, and anablated no-gate variant, PSO-LSTM achieves statistically significant improvements on FirstNode Dead (FND) (Δ ≥ +4.68%, padj ≤ 0.023, δ ≥ +0.761), Last Node Dead (LND)(Δ ≥ +6.21%, padj ≤ 0.001, δ = +1.000), and cumulative throughput (Δ ≥ +2.50%,padj ≤ 0.001, δ = +1.000), while introducing no statistically significant Half Node Dead(HND) degradation (padj ≥ 0.088), confirming the gate’s non-disruptiveness during stablenetwork operation. Ablation analysis confirms that the reliability gate is the key mechanismdriving these improvements.