A Neuro-Fuzzy Digital Twin for Interpretable Cardiac Disease Recognition

We present a neuro-fuzzy digital twin for cardiac disease recognition on the PTB-XL dataset that keeps the accuracy of a strong convolutional model while exposing its reasoning as readable fuzzy rules. The key design choice is to separate the two jobs instead of forcing one network to do both. A convolutional backbone (xresnet1d101 with cross-lead attention) is trained first and acts as the predictor. A second phase then freezes the backbone and trains a five-expert mixture of interval Type-2 Adaptive Neuro-Fuzzy Inference Systems as an interpretation layer, attached through a learnable trust gate. Four experts read complementary signal streams (the raw 12-lead waveform, a Fourier–Bessel Series Expansion, a Tunable-Q Wavelet Transform, and a morphogram); a fifth reads 22 clinical and demographic descriptors. A concept bottleneck maps the backbone embedding to eight named clinical concepts (ST elevation, T inversion, QT prolongation, Sokolov–Lyon and Cornell voltages, RVH, axis deviation, QRS widening), so the downstream rules read in clinical language. Two genetic-algorithm stages keep the model compact: NSGA-II selects the clinical features, and a multi-objective rule-pruning search trades macro-F1 and hypertrophy recall against the number of active rules. On the official test fold the twin reaches a macro-F1 of 0.751 (bootstrap 95% CI 0.740 to 0.767), a macro-AUC of 0.928, a Matthews correlation coefficient of 0.672, and a macro expected calibration error of 0.021, which is above the Strodthoff xresnet1d101 baseline of 0.74 and well below the calibration error of our earlier coupled design. Because the fuzzy layer is attached as a trust-gated residual and the best epoch is chosen on threshold-optimized validation F1, the interpretable twin matches the backbone within about 0.004 (twin 0.751 versus backbone 0.754) rather than paying the usual “interpretability tax”. The fuzzy mixture on its own still reaches a macro-F1 of 0.69, so the rules carry real diagnostic signal. Subgroup macro-F1 stays within about 0.045 across sex, age, and body-mass groups. A concept-level intervention simulator estimates how cardioactive drug classes would shift a patient’s risk by perturbing the named concepts and re-running the same twin.