{"id":"aead06c2-590b-4a99-81c6-d2732827dac0","arxiv_id":"2608.13425","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Frozen self-supervised speech representations transfer poorly across languages and corpora; Parkinson's-trained classifiers treat dementia speech like Parkinson's, indicating corpus-driven rather than disease-specific signal.","lead":"This paper tested whether speech-based AI models that detect Parkinson's disease still work under new languages, recording conditions, and patient groups, and found that the useful signal survives mostly because it comes from the dataset rather than from the disease. The results warn that these models can mistake dementia speech for Parkinson's speech, so clinical use needs much stricter validation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corpus–language confound: 'corpus-driven' attribution in Sections III-A and IV is untested because DE/ES/CZ vary jointly; no same-language control for layer selection.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: source corpus is conflated with language and recording conditions. This is the most load-bearing issue because the paper's headline answer to the title question ('Corpus') rests directly on that attribution. The TREND non-specificity result (Section III-E) is less affected, since it tests PD vs dementia within a single target cohort, but the 'primarily driven by the corpus' claim in the abstract and Discussion is not supported without a language-controlled comparison. The proposed ES/ES-e test is feasible with existing data (both subsets are described in Section II-A and used in S2), and it would settle whether the observed layer-selection differences are due to corpus/condition or language. Because the concern is addressable by additional analysis and the authors already interpret their specificity result conservatively ('absence of evidence, not evidence of absence'), the verdict remains CONDITIONAL; no adjustment is needed beyond the reader's recommendation.","tokens_in":12376,"tokens_out":12177,"duration_ms":123698,"concrete_test":"Recompute REF layer selection for the ES and ES-e subsets (same Spanish language, different recording condition; Section II-A) for DDK and READ across all nine backbones, using the same 10 seeds × 5-fold protocol. If ES vs ES-e argmax layers differ by more than 3 normalized-layer units at fixed backbone, corpus/recording condition drives selection independently of language; if they coincide, the DE/ES/CZ differences are more plausibly language-driven, and the 'corpus-driven' claim should be revised to 'language-driven' or 'dataset-driven'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that 'what survives across transfer settings is primarily driven by the corpus' (Discussion) depends on attributing layer-selection and transfer differences to corpus identity rather than to language or recording conditions. The three source corpora (DE, ES, CZ) differ simultaneously in language, recording hardware, and protocol (Table I), and no within-language cross-corpus comparison is reported. The only same-language pair, ES and ES-e, is used solely for S2 (+Condition) transfer (Section III-C) and is never analyzed for layer selection. Consequently, the Section III-A finding that the optimal layer is 'determined primarily by the source dataset' cannot distinguish corpus from language. If language is the operative factor, the title's answer changes from 'Corpus' to 'Language', and the practical implication (e.g., whether language-matched transfer would succeed) changes. The issue is compounded by the fact that the argmax layer is a point estimate: Table III reports cross-corpus σ of normalized depth but no within-corpus seed variance, so the large σ values (e.g., WavLM-L DDK σ=0.43) could reflect sampling noise rather than a robust corpus effect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether self-supervised speech representations capture Parkinson's disease (PD) specific signal that survives cross-lingual transfer. Using nine SSL backbones plus handcrafted eGeMAPS features, a logistic regression probe, and three special-session corpora (Spanish, German, Czech) together with an independent German TREND cohort, the authors define five transfer scenarios of increasing distribution shift. They report two main findings: (1) the optimal SSL layer in within-corpus reference evaluation varies substantially across source corpora, suggesting corpus-dependent layer selection; and (2) classifiers transferred to TREND separate PD from healthy controls but fail to separate PD from dementia, indicating a lack of pathological specificity. The paper interprets these results as evidence that transfer performance is driven by corpus-specific factors rather than a robust PD-specific speech signature, while carefully noting that this is absence of evidence, not evidence of absence.","tokens_in":12597,"tokens_out":5810,"duration_ms":54765,"significance":"If fully supported, the findings would make a valuable contribution to the debate on what speech-based PD classifiers actually learn, and the paper is among the first to evaluate cross-pathology specificity in this setting. The experimental design is thoughtful: multiple SSL backbones spanning capacity, fine-tuning, and pretraining language; layer selection is isolated from transfer evaluation; and the held-out TREND cohort provides a clinically meaningful cross-disease comparison. The authors also explicitly avoid overclaiming by framing the null result as absence of evidence, not evidence of absence. However, the central attribution of the findings to 'corpus' is currently underdetermined by the design, because the source corpora vary jointly in language and recording conditions, and the layer-selection results lack quantification of the sampling variability of the argmax layer.","major_comments":[{"comment":"The conclusion that layer selection is 'corpus-dependent' conflates corpus identity with language and recording conditions: the three source corpora (DE, ES, CZ) differ simultaneously in language, recording hardware, and protocol (Table I), and no within-language cross-corpus comparison is provided for layer selection. The only same-language pair (ES and ES-e) is used exclusively in Scenario 2 and is never analyzed for layer selection. As a result, the observed differences in optimal layer could equally be attributed to language or recording condition, and the title's answer 'Corpus' is not supported over 'Language' or 'Recording condition'. The authors should either add a same-language control (e.g., compare layer selection between ES and ES-e, noting the task limitations) or temper the conclusion to 'dataset identity' with an explicit acknowledgment of this confound.","section":"Section III-A, Table III and Figure 2"},{"comment":"The argmax layer is reported as a single number per corpus-backbone-task, and the σ values in Table III quantify variation across corpora, not within-corpus variability across the 10 seeds and folds. Since the layer-wise balanced-accuracy curves in Figure 2 are often flat (especially for ES), the argmax estimate may be dominated by sampling noise. Provide bootstrap or permutation confidence intervals for the selected layer per corpus, or a statistical test (e.g., a corpus-by-layer interaction in balanced accuracy) to show that cross-corpus differences in optimal depth are larger than within-corpus seed variation. Without this, large σ values such as WavLM-L on DDK (0.43) cannot be interpreted as a robust corpus effect.","section":"Section III-A, Table III"},{"comment":"The non-specificity claim is evaluated only on 9 of 540 transfer combinations, selected by thresholds on target-data AUC and BA greater than 0.6. This selection uses target labels to choose which combinations to interpret, and the authors do not report how many of the remaining 531 combinations would have shown PD-versus-dementia separation had they exceeded the threshold. Consequently, the second main finding may depend on a small and potentially unrepresentative subset. Please report the distribution of PD-versus-dementia separation (e.g., AUC or Mann–Whitney U p-values) across all 540 combinations, or justify why the threshold does not qualitatively affect the conclusion.","section":"Section III-E"},{"comment":"The description of the residualisation analysis is too brief to verify: 'We than adjusted the PD probability scores for these covariates using leave-one-subject-out (LOSO) regression. The residualised performance drops to chance level (AUC = 0.50–0.55)'. It is unclear whether the residuals are computed on predicted probabilities or logits, which TREND groups are included in the residualized evaluation, and what the residualized AUC is for PD versus HC(PD) as opposed to PD versus dementia. Please clarify the procedure and report the full results, including the demographic-only model's performance on the same groups.","section":"Section III-E, demographics adjustment"}],"minor_comments":[{"comment":"The sentence fragment 'Identity-preserving re-takes. S1 (+Re-Take) have a small degradation of performance' should be rewritten, for example as 'Identity-preserving re-takes (S1) show only a small degradation in balanced accuracy (−1.9 BA points).'","section":"Section IV, Discussion"},{"comment":"The phrase 'We than adjusted' should be 'We then adjusted'.","section":"Section III-E"},{"comment":"The caption text 'n= 2000test-set resamples' is missing a space and should read 'n = 2000 test-set resamples'.","section":"Figures 3 and 4 captions"},{"comment":"The sentence 'These trends are broadly consistent across backbone variance axis' should be 'across backbone variance axes'.","section":"Section III-A"},{"comment":"The claim that 'no consistent advantage of multilingual (Mu) over monolingual (Mo) SSL backbones is observed' is made without supporting statistical tests or effect sizes; adding a small number of tests or confidence intervals would strengthen the claim.","section":"Section III-D"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and addresses a timely question. The main risk is the corpus–language–recording confound: the title's answer is 'Corpus', but the design cannot distinguish corpus from language or recording condition for the layer-selection result. This is fixable by adding a within-language control (e.g., ES vs ES-e layer analysis) or by softening the conclusion; either would substantially change the paper's central claim. The lack of significance testing on the argmax layer is also a concrete, addressable gap. If the authors respond to these two points convincingly, the paper could be a valuable contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper's most valuable result is the cross-disease check: classifiers trained to separate PD from healthy controls transfer to the German TREND cohort and cannot separate PD from dementia even when they can separate both patient groups from controls. That is a useful caution for the speech-biomarker field, and the demographic residualization makes it more credible. Second, the headline answer 'corpus' is not actually established. The three training corpora differ simultaneously in language, recording hardware, and protocol, and the only same-language pair (ES / ES-e) is used only for recording-condition transfer, never for layer selection. So the claim that the optimal SSL layer is 'determined primarily by the source dataset' cannot distinguish corpus identity from language. If language is the operative factor, the title's answer changes. Same-language, cross-corpus layer-selection evidence would settle it.\n\nWhat is genuinely new: the five-scenario progressive-shift framework is a clean way to separate re-take, condition, language, task, and pathology shifts; the layer-wise evaluation across nine SSL backbones is thorough; and the PD-vs-dementia specificity test looks new relative to the cited literature. The authors also phrase the null conservatively—'absence of evidence, not evidence of absence'—which is the right register.\n\nSoft spots, in proportion. The corpus–language confound is the main one, and it affects the interpretation of Section III-A and the Discussion, not the dementia null result, because the German-only transfer (S4) already fails specificity. The argmax layer selection is a point estimate; Table III reports variation across corpora but no within-corpus seed variance, so the large sigma values could partly be sampling noise. Minor. The specificity analysis is restricted to 9 of 540 transfer combinations chosen by BA/AUC thresholds—post hoc, but defensible, because asking about specificity is only meaningful when there is baseline separation; still, the 'consistent across all 9' claim rests on a small, selected set. Small TREND cohorts (36 dementia, 18 PD) are acknowledged. Citation pattern is fine: the authors cite prior evidence that SSL embeddings cluster by corpus and do not overstate novelty.\n\nMy bottom line: the cautionary result is real, the interpretation is overreached at the margins. The paper deserves a serious referee. I would send it with a request for a same-language corpus analysis or a clear admission that the corpus/language distinction is unresolved, plus uncertainty quantification for layer selection.","headline":"The PD-vs-dementia specificity result is the real contribution and broadly holds; the 'corpus-driven' headline is plausible but the design cannot separate corpus from language.","tokens_in":13129,"tokens_out":2479,"would_cite":true,"duration_ms":25345,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Speech-based Parkinson's detection models trained on frozen self-supervised embeddings learn a corpus-driven, disorder-general signal rather than a PD-specific signature, and under cross-lingual transfer they fail to distinguish…","keywords":["Parkinson's disease","self-supervised speech representations","cross-lingual transfer","dementia","speech biomarkers","layer-wise analysis","diagnostic specificity","corpus confounds"],"falsifier":"A within-language cross-corpus transfer experiment, such as training on one German PD corpus and testing on another German PD corpus recorded under different conditions, that preserved PD-versus-dementia separation and stable layer selection would contradict the corpus-driven interpretation; alternatively, finding a target pathology (for example depression or essential tremor) for which the transferred PD classifiers do not show elevated probabilities would falsify the disorder-general claim.","tokens_in":12152,"feed_emoji":"🧠","tokens_out":4349,"duration_ms":41615,"temperature":0.7,"pith_summary":"This paper asks whether self-supervised speech representations used for Parkinson's disease (PD) detection capture disease-specific motor or cognitive characteristics, or instead exploit dataset-specific cues. The authors train logistic-regression probes on frozen SSL embeddings from nine backbones across Spanish, German, and Czech PD corpora, then transfer the resulting classifiers to an independent German cohort that includes both PD and dementia. They report two findings: the best-performing SSL layer is determined mainly by the source corpus rather than the architecture, and classifiers that separate PD from healthy controls assign similarly high PD probabilities to dementia speech. The authors conclude that what survives cross-lingual transfer is a generic divergence from healthy speech, not a robust PD-specific signature, and they frame this as absence of evidence for PD specificity with direct implications for clinical deployment.","feed_headline":"PD speech AI transfers corpus cues, not a PD signature","feed_subtitle":"Models trained for Parkinson's separate patients from healthy speakers but cannot pick out Parkinson's from dementia.","key_machinery":"The central machinery is a five-scenario transfer evaluation built on frozen SSL speech encoders and a low-capacity logistic-regression probe. For each corpus, backbone, and task, the representation layer with the highest mean balanced accuracy is selected under within-corpus 5-fold cross-validation (REF) and then fixed across all transfer scenarios, so that performance differences reflect distribution shift rather than layer choice. The load-bearing comparison is the transfer of PD-trained classifiers to the TREND cohort, which contains both PD and dementia participants, allowing the authors to test whether the transferred signal is specific to Parkinson's or merely indicates a general divergence from healthy speech.","core_discovery":"On the paper's own terms, the central discovery is that the discriminative signal in frozen SSL speech representations that transfers across corpora is corpus-driven and disorder-general. Under five progressively harder distribution shifts, covering re-takes, recording conditions, language, task, and combined language-task shifts, PD-versus-healthy balanced accuracy degrades steadily, from a mean change of −1.9 points for re-takes to −16.3 points for cross-lingual transfer. In the clinically most relevant setting, transferring PD-trained classifiers to the TREND cohort, only 9 of 540 transfer combinations achieve at least moderate PD-versus-control separation (AUC and balanced accuracy above 0.6). Those 9 classifiers separate PD from matched healthy controls but fail to distinguish PD from dementia, and after adjusting for age, sex, and education the PD-versus-dementia discrimination drops to chance. The authors interpret this as evidence that the models preserve a general patient-healthy separation rather than a disease-specific structure.","pith_inferences":["A natural test of the corpus-driven interpretation would be a within-language cross-corpus transfer, such as training on one German PD corpus and testing on another German PD corpus recorded with different hardware, because the current design varies language, recording hardware, and protocol jointly with corpus.","The similar probabilities assigned to PD and dementia speech suggest the learned signal could be repurposed as a general marker of neurodegeneration or cognitive-motor decline, but this would require validation against several additional neurological conditions.","Because the non-specificity finding rests on only 9 transfer combinations that passed the performance threshold, the claim is strongest for those configurations; testing more target pathologies and more task combinations would clarify how broadly the disorder-general signal holds.","The instability of optimal layer choice across corpora could itself serve as a diagnostic of dataset shift: if the optimal layer depends mainly on the corpus, then a stable disease-specific speech signature is unlikely to be captured by any single SSL layer."],"forward_implications":["Speech-based PD classifiers with strong within-corpus accuracy can still lack diagnostic specificity, so evaluations that compare PD only against healthy controls are insufficient for clinical claims.","The optimal SSL layer for PD detection should not be assumed to transfer across corpora; layer choice varies substantially with the source dataset, especially for large backbones.","Multilingual pretraining (XLS-R, MMS) offers no consistent advantage over monolingual SSL backbones for cross-lingual PD detection.","Recording-condition mismatch degrades performance asymmetrically: training on noisy recordings transfers better to clean recordings than the reverse.","Performance degrades monotonically as distribution shift increases, with mean balanced-accuracy changes of −1.9 (re-take), −12.5 (recording condition), and −16.3 (cross-lingual) relative to the within-corpus reference."],"supporting_citations":[{"why":"Supplies the three special-session PD corpora and the prior cross-lingual dysarthria adaptation setting this study extends.","marker":"[17]"},{"why":"Provides the Czech PD corpus used as one of the three training sources.","marker":"[18]"},{"why":"Provides the Spanish PD corpus and its noisy second subset used for the recording-condition scenario.","marker":"[19]"},{"why":"Provides the German PD corpus that anchors the cross-task transfer scenario because it shares language with the TREND cohort.","marker":"[20]"},{"why":"Provides the TREND cohort with dementia and PD participants that makes the specificity analysis possible.","marker":"[21]"},{"why":"Defines the HuBERT backbone whose base and large variants are analysed layer-wise.","marker":"[27]"},{"why":"Defines the WavLM backbone whose base and large variants are analysed layer-wise.","marker":"[28]"},{"why":"Defines the wav2vec 2.0 backbone and its ASR-fine-tuned variant.","marker":"[29]"},{"why":"Provides the multilingual XLS-R backbone used to test whether multilingual pretraining helps cross-lingual transfer.","marker":"[30]"},{"why":"Provides the MMS backbone trained on over 1,000 languages, the broadest multilingual comparison point.","marker":"[31]"}],"fun_headline_variants":["Speech AI for PD is corpus-tuned, not disease-specific","Cross-lingual PD speech models learn corpus, not pathology","Parkinson speech AI: transfers generalize, not discriminate","What transfers in PD speech? Corpus, not PD signature","PD speech AI can't tell PD from dementia in transfer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion that transfer behaviour is driven by the corpus assumes that the source corpus can be separated from language and recording conditions, since the three training corpora differ jointly in language, recording hardware, and protocol with no within-language cross-corpus control.","fun_headline_variants_meta":{"raw":{"variants":["Speech AI for PD is corpus-tuned, not disease-specific","Cross-lingual PD speech models learn corpus, not pathology","Parkinson speech AI: transfers generalize, not discriminate","What transfers in PD speech? Corpus, not PD signature","PD speech AI can't tell PD from dementia in transfer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000795,"raw_usage":{"total_tokens":3487,"prompt_tokens":921,"completion_tokens":2566,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2484}},"tokens_in":537,"tokens_out":2566,"duration_ms":17004,"temperature":1.0,"reasoning_tokens":2484,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:16:52.514913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A within-language cross-corpus transfer experiment, such as training on one German PD corpus and testing on another German PD corpus recorded under different conditions, that preserved PD-versus-dementia separation and stable layer selection would contradict the corpus-driven interpretation; alternatively, finding a target pathology (for example depression or essential tremor) for which the transferred PD classifiers do not show elevated probabilities would falsify the disorder-general claim.","supporting_citations":[{"cited_title":"Multi-Modal Decentralized Hybrid Learning for Early Parkinson’s Detection Using V oice Biomarkers and Contrastive Speech Embeddings,","cited_arxiv_id":null,"evidence_quote":"Provides the Czech PD corpus used as one of the three training sources."},{"cited_title":"Emotional State Modeling for the Assessment of Depression in Parkinson’s Disease,","cited_arxiv_id":null,"evidence_quote":"Provides the Spanish PD corpus and its noisy second subset used for the recording-condition scenario."},{"cited_title":"Automatic speech-based assessment to discriminate Parkinson’s disease from essential tremor with a cross-language approach,","cited_arxiv_id":null,"evidence_quote":"Provides the German PD corpus that anchors the cross-task transfer scenario because it shares language with the TREND cohort."},{"cited_title":"(2026) T ¨ubinger Erhebung von Risikofaktoren zur Erkennung von Neurodegeneration (TREND)","cited_arxiv_id":null,"evidence_quote":"Provides the TREND cohort with dementia and PD participants that makes the specificity analysis possible."},{"cited_title":"openSMILE: the Munich versatile and fast open-source audio feature extractor,","cited_arxiv_id":null,"evidence_quote":"Defines the WavLM backbone whose base and large variants are analysed layer-wise."},{"cited_title":"WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing,","cited_arxiv_id":null,"evidence_quote":"Provides the multilingual XLS-R backbone used to test whether multilingual pretraining helps cross-lingual transfer."}],"review_version":1}