What Survives Phonocardiogram Tokenization? Tokenizer Design and Interface Geometry Shape Clinical Recoverability

Abstract. Discrete tokenizers offer compact, reusable interfaces between physiological signals and sequence foundation models, as in neural audio codecs and tokenized time-series models [1,2], but clinically relevant information may be lost during continuous representation formation, quantization, or downstream token embedding. Treating these stages as one bottleneck obscures where fixes belong. We localize loss in 3,163 phonocardiogram recordings from 942 subjects in the public CirCor DigiScope cohort [3], using patient-disjoint splits. Held-out murmur evaluation included 462 recordings from 132 subjects (78 murmur-positive); recording-site evaluation included 485 recordings from 142 subjects. A fixed bidirectional-GRU probe measures recoverability 𝑅, the normalized reduction in held-out log loss relative to label entropy, using recording-level predictions and patient-clustered bootstrap intervals. At comparable medium marginal entropy rates, a wave form convolutional finite-scalar-quantization (FSQ) tokenizer [4] at 265 bits/s showed pre-quantization collapse: murmur 𝑅 fell from 0.535 on matched log-spectral inputs to 0.025 in continuous latents and 0.008 after quantization. Across three preregistered learned-codebook EMA-VQ seeds [5] (two meeting all localization criteria), raw-to-latent losses were 0.320–0.541, where as post-minus-pre effects were −0.032 to +0.059. Multi-resolution spectral reconstruction improved token murmur recoverability in all three paired FSQ seeds (mean Δ𝑅 = 0.092), although rate rose 1.7%–6.2% and two of three absolute-𝑅 intervals crossed zero. In contrast, an AST-based spectral Transformer [6] at 283 bits/s preserved recoverability from input (𝑅 = 0.544) through continuous latents (0.539) and numeric FSQ coordinates (0.538); at 143 bits/s, rounding reduced 𝑅 from 0.534 to 0.489. Because the tokenizer packages also differed in input, decoder, objective, and optimization, these results establish design dependence rather than an architecture ranking. Across Transformer rates of 562, 283, and 143 bits/s (a 3.9-fold reduction), a factorized categorical interface retained murmur 𝑅 = 0.525, 0.529, and 0.500 (AUROC = 0.941, 0.945, and 0.926). Encoding the identical FSQ states as unrelated flat IDs gave 𝑅 = 0.402, 0.417, and 0.342 (AUROC = 0.880, 0.873, and 0.842). Under factorization, the mean high-to-low decline was 0.026, and no seed exceeded the preregistered 0.10 threshold. The medium-rate interface gap (Δ𝑅 = 0.112) was therefore 4.3 times this rate effect, despite a one-to-one mapping and only 368 interface parameters versus 64,000 for flat IDs. With more than 99.999% held-out token coverage, this is an accessibility effect, not information deletion. Natural factorization beat five matched random factorizations for each of three tokenizer seeds (median paired Δ𝑅 = 0.141–0.197) and improved recording-site 𝑅 by 0.054–0.089 across all nine rate-seed combinations. Clinical recoverability therefore depends on two distinct bottlenecks: tokenizer training determines what survives compression, while interface geometry determines how readily a bounded downstream model can access it. Physiological foundation-model pipelines should audit both and preserve quantizer factor structure. These experiments do not establish a causal architecture ranking or cross-dataset generality.