Rate-Based Cardiac Rhythm Triage

Pooled confusion matrix — AAMI DS2, hybrid GT

The Irregular-class fix. v0.5 reported 0.84 Irregular recall against a CV-based ground truth that was partially circular with the detector's own CV rule. Under the more honest clinician-annotated hybrid GT, the v0.5 detector dropped to 0.19 Irregular recall — AFIB segments with stable short RR were silently misclassified. v0.6 adds range/RR > 0.40 to the classifier and lifts Irregular recall to 0.78 on the same data. Trade-off: Normal / Tachy / Brady recalls drop ~10–30 pp because the new rule fires more often. Stable rapid AFIB cannot be reliably distinguished from sinus tachy by short-window RR features alone — that is a fundamental limit of rate-regime triage and is not solved here.

Per-record accuracy — honest patient-level variability

Range: 86.5 % (record 233, frequent PVCs & AF) ↔ 100.0 % (records 100, 103, 111). Accuracy drops on records with frequent ventricular ectopics (200, 214, 233) and AF (232) because rate-regime classification is fundamentally limited when ectopic beats disrupt the RR pattern. AAMI N/S/V/F/Q columns are traceability only: CricketBrain does not classify morphology classes.

Baseline comparison — CricketBrain on par with a hand-coded rule

Honest finding (under hybrid clinician GT). A 50-line band-gate + RR-window rule (ThresholdBurstBaseline) is ~0.3 pp better than CricketBrain on AAMI DS2. The trivial 1-second-window rule fails badly at 18.2 %, so the task is non-trivial — but the neuromorphic core does not beat a careful classical rule on rate-regime triage. CricketBrain's value proposition is: 928 bytes RAM, deterministic, zero training, the same circuit reused across UC02 bearings / UC03 marine / UC04 grid, not "we solve a hard problem".

Drift-robustness sweep — honest result

We hypothesised CricketBrain's Gaussian tuning would degrade more gracefully than the rule's hard ±10 % band-pass under QRS carrier-frequency drift. The data did not support that hypothesis. A synthetic drift sweep (15 offsets in [−20 %, +20 %]) shows the rule has clean cliffs at ±10 % (its band edge), while CricketBrain's response is non-monotonic: it works at off ∈ {0 %, ±1 %, ±5 %, +20 %} and emits zero classifications at off ∈ {−3 %, −7 %, −10 %, +3 %, +7 %, +10 %, ±15 %}. Both systems break around the same drift band, but neither is meaningfully more robust. CSV at results/cardiac_drift_sweep.csv.

d′ / TPR / FPR (Green & Swets 1966)

200 signal + 200 noise trials per condition on synthetic P-QRS-T cycles. Wilson 95 % confidence intervals. d′ uses the Hautus 1995 log-linear correction. This SDT setup measures synthetic-rhythm discriminability, not real-patient performance — for that, see the v0.4 MIT-BIH section above.

Noise Rejection: Without vs. With Preprocessor

The EcgPreprocessor applies temporal consistency filtering: in-band signals must persist for 3+ consecutive steps to pass through. Single-step noise spikes are rejected. Gap tolerance (2 steps) bridges brief interruptions within real QRS bursts.

Other Stress Tests

The 5-neuron Münster circuit

CricketBrain's core is a hardwired 5-neuron / 6-synapse delay-line coincidence detector inspired by the field cricket's auditory T-cell circuit (Münster model). Two delay lines feed a coincidence neuron: only when both inputs arrive within the integration window does it fire. There is no matrix multiplication, no weights to train, no GPU — just timing.

Why this matters for UC01. The same 5-neuron circuit is reused across UC02 (bearing vibrations), UC03 (marine acoustics) and UC04 (power-grid harmonics) by changing only min_freq/max_freq. No retraining, no new architecture — the “Daseinsberechtigung” is cross-domain reuse in 928 bytes, not state-of-the-art rate-regime accuracy on any single task. On AAMI DS2 a 50-line band-gate rule is on par with this circuit; what the circuit buys you is portability across signal domains and zero training data.

Cardiac detection pipeline

RAM, Model Size & Latency

Accuracy & Training Data

CricketBrain's 78.4 % pooled accuracy on real MIT-BIH AAMI DS2 with clinician rhythm-annotation ground truth is on rate-regime triage only — and a hand-coded band-gate + RR-window rule reaches 78.7 % on the same data, so the neuromorphic core does not beat classical rules here. Morphology-aware CNNs (Tiny MF-CNN, Stanford DNN) tackle the harder 5- and 12-class problems with 15 KB – 40 MB RAM and labelled training data. Different problems, different operating envelopes — see docs/competitive_analysis.md.

When to Pick Which

Full sourced breakdown (Nuzzo 2023, Hannun 2019, Apple FDA DEN180044): docs/competitive_analysis.md.