Preprocessing Filter Choice Affects Severity Calibration but Not Accuracy in Cardiac Murmur Classification

Phonocardiogram (PCG) preprocessing pipelines for automated cardiac auscultation vary widely in their choice of denoising filter, yet the downstream effect of that choice on murmur severity classification is rarely tested directly. We report a pre-registered factorial comparison of three preprocessing conditions, a causal Butterworth bandpass, an autoregressive Kalman filter, and a minimal lowpass restriction, crossed with two model families, a squeeze-excitation residual CNN and a bidirectional LSTM, on the public portion of the CirCor DigiScope dataset. Using a curated sample of 120 murmur-positive patients (Levine grades I/VI, II/VI, III/VI+) and 10-fold patient-level cross-validation (60 trained models), we find no significant effect of filter choice on macro-F1 or balanced accuracy, and no significant Filter $\times$ Architecture interaction, under either a three-way or a collapsed two-way severity scheme. The achieved minimum detectable effect (Cohen's $d\approx0.91$) bounds this as a null over large effects rather than an unqualified absence. A finer-grained ordinal analysis nonetheless detects a small, directionally consistent bias: minimal filtering shifts predictions toward higher severity relative to Butterworth, replicated across both label schemes. Unexpectedly, model architecture had a significant, FDR-surviving effect larger than any filter effect. For low-resource screening in African clinical settings, this suggests classification accuracy is robust to the exact denoising filter, but ordinal calibration is not, which matters directly for severity-based triage thresholds.