{"id":"fa62b881-ace1-4a11-9837-e65953631b7a","arxiv_id":"2505.20731","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SCORE, a Poisson-adapted latent factor mixture model with hybrid EM and Gaussian variational inference, yields consistent semi-supervised patient embeddings and better MS disability phenotyping than existing methods.","lead":"The paper introduces SCORE, a semi-supervised statistical framework that learns patient clusters and embeddings from high-dimensional electronic health records by combining a Poisson mixture model with pre-trained code representations. It provides convergence guarantees and shows improved disability phenotyping for multiple sclerosis patients when labels are scarce.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The practical and theoretical claims both hinge on V spanning the true latent subspace; the paper's misspecification simulations never perturb V, so the most load-bearing condition is untested.","rationale":"The reader's verdict (CONDITIONAL, weakest assumption on V) is well aligned. The paper is a well-motivated methodological contribution with a plausible theoretical framework and encouraging simulations; the central claim is internally consistent under its assumptions. But the single place where the argument has real exposure is Assumption 2(i): the entire machinery treats V as the true latent span, and this is also exactly the element that is estimated externally in the application. The omitted proofs of Lemma 1 and Theorem 2 are a separate verification gap that supports the CONDITIONAL verdict rather than full acceptance, but they are not a specific mathematical flaw. The V concern is concrete and testable: the current misspecification simulations are too narrow to catch it, and the Discussion's own admission shows the authors recognize the limitation. If the proposed simulation shows severe degradation, the claim that unlabeled data and pre-trained embeddings deliver label-scarce robustness would need substantial qualification; if it shows only mild sensitivity, the concern is resolved. I therefore recommend no change to the reader's CONDITIONAL verdict: the paper should be accepted only after V-misalignment robustness is assessed and the missing proofs are supplied.","tokens_in":858,"tokens_out":808,"duration_ms":243322,"concrete_test":"Run the Section 4 simulation in the regime n=50, N=5000, p=400, q=20 with V_used equal to the first q columns of V_true, but generate the true rates as Z = V_used (B D + W) + V_perp eta, where V_perp is orthogonal to V_used and eta is a normal vector with variance sigma^2. Sweep sigma^2 from 0 to a level comparable to Var(V_used W), and also replace V_used by embeddings estimated from an independent training cohort. If semi-supervised SCORE's relative F-norm error or AUC degrades sharply with sigma^2, or if its advantage over the supervised estimator disappears, the n-free guarantee and the transfer-learning claim are not robust to V misalignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 2(i) fixes the pre-trained embedding matrix V as the exact span of the PALM latent factors. All of Theorem 2, Corollary 1, and the transfer-learning benefit flow from this: SCORE reduces parameters by replacing the unknown loading matrix with V, and the n-free N^{-1/2} rate is derived under this model. In the EHR application, V is obtained as the SVD of external multi-source concept embeddings (plus five hand-selected disability features), with no guarantee that the target MS population's latent disease structure lies in that span. The misspecification simulations in Section 4 only add an unobserved factor C_i inside the same V column space; they never rotate, truncate, or contaminate V so that the true Z has an orthogonal component. The Discussion concedes that when external and target structures diverge substantially, rescaling eigenvalues is insufficient. Therefore the central practical claim, that pre-trained embeddings make SCORE robust to label scarcity, is least secure exactly where V is misaligned, and no experiment or theorem in the manuscript bounds this failure mode. The omitted proofs of Lemma 1 and Theorem 2 compound this: the contraction argument that supports the n-free rate is not independently verifiable in the current version.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCORE, a semi-supervised representation-learning framework for high-dimensional EHR count data. The model is a Poisson-adapted latent factor mixture (PALM) in which a pre-trained code-embedding matrix V is used as a known loading subspace, and only the coefficient matrix B, the random-effect covariance Λ, and the label model b are estimated. Estimation is carried out by a hybrid EM algorithm that uses a supervised ELBO-based estimator for initialization and Gaussian variational approximations for the unlabeled subjects. The main theoretical results are a linear-contraction lemma for the EM-GVA iterates (Lemma 1), a convergence rate for the semi-supervised estimator that is free of the labeled sample size n (Theorem 2), and a rate for the predicted phenotype probability and patient embedding for a new subject (Corollary 1). The empirical section reports simulation studies and an application to multiple sclerosis disability phenotyping in 16,091 patients, claiming improved AUC, PRAUC, and Brier score over several baselines, especially under label scarcity.","tokens_in":20126,"tokens_out":3018,"duration_ms":34565,"significance":"If the theoretical results hold, the paper would be a useful contribution: it extends the theory of Gaussian variational approximation to a low-rank Poisson mixture with diverging latent dimension and semi-supervised data, and it gives a concrete transfer-learning construction based on pre-trained EHR code embeddings. The explicit SNR condition and the n-free N^{-1/2} rate are interesting and would strengthen the case for using unlabeled EHR data in rare-disease settings. The simulation design with 200 replications and the real MS application are also valuable, and the paper is honest about the limitation that the external embedding structure may not match the target population. However, the manuscript as submitted cannot be fully verified because the proofs of the main theorems are deferred to an appendix that is not included, and the robustness analysis does not address the most threatening deviation from the key assumption that V spans the true latent subspace.","major_comments":[{"comment":"The main theoretical claims—the linear contraction of the EM-GVA algorithm and the n-free convergence rate of the SCORE estimator—are stated without proofs in the main text, and the appendix containing the proofs is not present in the submitted version. Since these results are the paper's central contribution, the manuscript currently does not allow independent verification of the derivation. The authors should either include the full proofs in an appendix or provide a detailed proof sketch in the main text, with the omitted technical lemmas clearly stated.","section":"Section 3, Theorem 2 and Lemma 1"},{"comment":"The misspecification simulations are constructed as Z_i = V( eB eD_i + W_i) with the same pre-trained V, so the additional unobserved factor C_i lies inside the column space of V. This does not test the case where the true latent structure has a component orthogonal to span(V), which is the principal threat to Assumption 2(i). The Discussion concedes that when the external and target factor structures diverge substantially, methods beyond rescaling eigenvalues may be needed. Because the practical claim that pre-trained embeddings make SCORE robust to label scarcity depends on this alignment, the manuscript should add simulations with a rotated, truncated, or contaminated V, or provide a formal analysis of the misalignment error.","section":"Section 4.1, misspecification setting (e), and Section 6"},{"comment":"The empirical comparisons are not fully reproducible from the information given. The manuscript does not specify the values or tuning of K_B and K_M, the number of EM iterations T, learning rates, convergence criteria, or the choice of q in the application. For the benchmark methods (Lasso, random forest, XGBoost, NN(sup), NN(VAE)), the tuning procedure, network architecture, and training details are omitted. Since the paper claims superior finite-sample performance, these implementation details should be reported or the code should be made available.","section":"Sections 4.1 and 5"}],"minor_comments":[{"comment":"The notation for A_{ij}^{(y)} is introduced as E_{W~φ_i^{(y)}} exp{V_j^T(B U_{iy} + W)}, but the superscript y is dropped in parts of the displayed expression; please make the dependence on y explicit throughout the equation and its surrounding text.","section":"Section 2.2, Eq. (3)"},{"comment":"There are minor typos: 'Chique2018' in Table 1 should be 'Chiquet et al. [2018]', and 'preresentation' should be 'representation' in the sentence about patient representation quality.","section":"Section 4.1, Table 1 and text"},{"comment":"The aggregation window of 6 months before and after each visit is a free parameter, but no sensitivity analysis is reported; please justify this choice or show that the results are not sensitive to it.","section":"Section 5"},{"comment":"The statement that allowing Λ to be freely estimated adjusts for 'scale heterogeneity as well as the rotation ambiguity' is informal; a precise statement of which aspects of the loading matrix are identifiable when V is fixed would help the reader understand the model's identification.","section":"Remark 1"}],"recommendation":"major_revision","confidential_remarks":"The absence of the appendix is a serious issue for a theory-heavy paper: the main theorems cannot be checked from the submitted materials. I would recommend asking the authors to provide the complete appendix before the review can be completed. The missing misalignment experiment is also important because the key transfer-learning benefit is exactly in the regime where V may not match the target population."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SCORE is a serious piece of methodology: it takes the Poisson log-normal mixture, adds a pre-trained column space V to cut parameters, and wraps it in a hybrid EM-GVA fit with supervised initialization. The genuinely new theoretical piece is the error analysis under diverging q, and the headline result that the semi-supervised estimator's rate is free of the labeled sample size n (Theorem 2) is meaningful if the assumptions hold. I also give credit for the mixture-of-GVAs posterior idea, which is more honest than a single Gaussian when latent clusters exist. The simulations and MS application are consistent with the theory and show real gains in label-scarce settings.\n\nThe soft spots are concentrated in one place, and the stress-test note has it right. Assumption 2(i) fixes V as the exact span of the true latent factors. Everything—the parameter reduction, the transfer learning, the n-free rate—flows from that. The misspecification simulations add an unobserved factor but keep it inside the V column space; they never rotate, truncate, or contaminate V so the truth has an orthogonal component. So the most load-bearing condition is empirically untested, and the Discussion concedes that rescaling eigenvalues may not be enough when external and target structures diverge. That concession is honest, but it means the practical claim is least secure in exactly the regime where the method is supposed to help most.\n\nOther issues are more mundane. The appendix with the proofs is referenced but not present in this version, so Lemma 1 and Theorem 2 are not independently verifiable. No code or data are shipped. Baseline implementation details are thin; I would want more on hyperparameters and the neural-network comparators before treating the real-data comparison as definitive. These are fixable, not fatal.\n\nThe paper is for methodologists working on EHR phenotyping and on variational inference for count factor models. It deserves a serious referee. My recommendation: send it to review, but condition acceptance on the full appendix and on a re-run of the misspecification simulations where V itself is perturbed—rotate it by a random orthogonal matrix, truncate it, or add a sparse orthogonal component—so the target structure has a component outside span(V). If the method still degrades gracefully, the central claim is much stronger. If it falls apart, the honest Discussion already tells you where the boundary is.","headline":"SCORE is a credible semi-supervised phenotyping method with a genuine diverging-q theoretical result, but the central transfer-learning claim rests on an untested assumption about V spanning the true latent space.","tokens_in":20641,"tokens_out":1869,"would_cite":true,"duration_ms":20498,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62H25","62F12","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"SCORE's error rate for EHR phenotyping stops depending on the number of labeled patients.","keywords":["factor mixture model","embedding","Gaussian variational approximation","Expectation-Maximization algorithm","EHR codified data","semi-supervised learning","patient phenotyping","representation learning"],"falsifier":"Run the PALM simulation with a known $V$, then rerun SCORE with $V$ altered by swapping one column for a random vector while holding $n$ fixed and doubling $N$; if the estimation error stops falling and stays above the predicted $\\psi q^{1/2}/N^{1/2}$ term, the span assumption is what carries the result.","tokens_in":19701,"feed_emoji":"🩺","tokens_out":13358,"duration_ms":119977,"temperature":0.7,"pith_summary":"The paper proposes SCORE, a semi-supervised framework that clusters patients and learns patient embeddings from sparse, high-dimensional EHR count data when labeled examples are scarce. It models the counts with a Poisson-adapted latent factor mixture (PALM) that treats pre-trained EHR code embeddings as a known linear subspace, so only scaling parameters, not the full loading matrix, must be estimated. The central theoretical claim is that a hybrid EM and Gaussian variational approximation algorithm, initialized from the supervised fit, converges at a rate dominated by the total sample size $N$ and the feature dimension $p$, with the labeled sample size $n$ absent from the leading term. If correct, label scarcity stops being the bottleneck: large unlabeled pools substitute for labels, and the learned embeddings transfer to related phenotypes such as EDSS-defined disability. Simulations and a multiple sclerosis EHR application support this conclusion.","feed_headline":"Unlabeled EHR data erase label scarcity in phenotyping","feed_subtitle":"SCORE's error rate tracks the total patient pool, not the labeled count, so rare-disease cohorts can be phenotyped.","key_machinery":"The engine is the Poisson-Adapted Latent factor Mixture (PALM): $X\\mid Z\\sim\\mathrm{Poisson}(e^Z)$ with conditionally independent coordinates, $Z=V(B\\tilde U_Y+W)$, $W\\sim N(0,\\Lambda)$, and $Y\\mid U\\sim\\mathrm{Bernoulli}(\\mathrm{expit}(b^T\\tilde U))$. The matrix $V\\in\\mathbb{R}^{p\\times q}$ is the fixed pre-trained embedding matrix of the $p$ EHR codes; treating its column space as known cuts the loading parameters from $pq$ to $q^2$ scaling parameters. The algorithm then uses a Gaussian variational approximation (GVA) with diagonal covariance per subject to obtain a closed-form ELBO, and the hybrid EM-GVA alternates between imputing the latent class probabilities $\\gamma^{\\mathrm{GVA}}$ for unlabeled subjects and maximizing the full-sample ELBO, initialized by the supervised ELBO estimator. The convergence proof runs through a linear-contraction lemma for the EM-GVA iterations, which requires the SNR $\\psi^2$ to exceed a constant and ties the final rate to $N$ and $p$ rather than to $n$.","core_discovery":"On the paper's own terms, the SCORE estimator recovers the PALM parameters $\\theta=(B,\\Lambda,b)$ at rate $$\\mathrm{Err}(\\hat{\\$\\theta$})=O_P(\\psi $q^{{1/2}}$/$N^{{1/2}}$+\\psi $q^{{2-\\varepsilon}}$/$p^{{1-\\varepsilon}}$),$$ which contains no labeled-sample-size term $n$, provided Assumptions 1-3 hold and the signal-to-noise ratio satisfies $\\psi^2>C_\\psi$ (Theorem 2). The estimator is built so that the supervised ELBO fit initializes the hybrid EM-GVA procedure and its influence decays geometrically, leaving only the unlabeled-data and GVA-approximation errors. For a new subject, the phenotyping probability $\\gamma_{\\mathrm{new}}^{\\mathrm{GVA}}$ converges to $\\Pr(Y=1\\mid \\bar\\xi_{\\mathrm{new}},U_{\\mathrm{new}})$ at rate $O_P(\\psi^2 q^{1/2}/N^{1/2}+\\psi q^{1/2-\\varepsilon}/p^{1/2-\\varepsilon})$, and the estimated embedding $\\hat E_{\\mathrm{new}}$ converges to the true latent embedding $\\bar\\xi_{\\mathrm{new}}$ (Corollary 1). Because Proposition 1 shows $\\Pr(Y=1\\mid\\bar\\xi,U)$ is at least as informative about $Y$ as the raw-count posterior, the learned low-dimensional representations are claimed to be the right object to use for downstream phenotyping and prediction.","pith_inferences":["Editorial extension: the $n$-free rate implies a scaling law for rare-disease studies - enlarge $N$ by pooling visits or collaborating centers and even very small label sets may be enough; the paper does not turn this into a recruitment rule.","Editorial extension: the span assumption can be stress-tested directly by deleting a column of $V$ aligned with the known outcome direction and seeing whether the GVA error floor $\\psi q^{2-\\varepsilon}/p^{1-\\varepsilon}$ remains as $N$ grows.","Editorial extension: the same contraction argument should carry to other exponential-family mixtures with known feature embeddings, such as negative binomial or zero-inflated counts, which the paper mentions only as future work.","Editorial extension: because unlabeled patients do not need an outcome, one learned embedding could serve several downstream labels simultaneously, so a single SCORE fit might phenotype multiple related conditions from one unlabeled pool."],"forward_implications":["Adding unlabeled patients shrinks parameter estimation error at a $\\psi q^{1/2}/N^{1/2}$ rate, so a cohort with only dozens of labeled patients can still be phenotyped when the unlabeled pool is large.","Only $O(\\log(N/n))$ EM iterations are needed to reach that rate, so the benefit of unlabeled data comes at a modest computational cost.","For new subjects, the predicted disability probability converges to $\\Pr(Y=1\\mid\\bar\\xi,U)$, which Proposition 1 shows is no less informative than using raw count features.","The model treats observed labels as noisy proxies for latent cluster membership, which supports cross-scale phenotyping: embeddings trained on PDDS-defined disability generalize to EDSS-defined disability and predict future progression.","Any consistent supervised estimator suffices for initialization, so the final error rate does not depend on which supervised baseline is used to start EM."],"supporting_citations":[{"why":"defines the multivariate Poisson-lognormal model that PALM extends to a latent-class mixture with covariates and known embeddings.","marker":"Aitchison and Ho [1989]"},{"why":"supplies the benchmark variational Poisson PCA method and the mean-field ELBO construction that SCORE adapts and extends to semi-supervised clustering.","marker":"Chiquet et al. [2018]"},{"why":"provides the two-step SVD baseline for item factor analysis that SCORE compares against and improves by modeling random effects and covariates.","marker":"Zhang et al. [2020]"},{"why":"gives the fixed-q joint maximum likelihood framework whose fixed-rank regime SCORE generalizes to diverging q.","marker":"Chen et al. [2019]"},{"why":"establishes GVA convergence for a rank-one Poisson mixed model, the result SCORE extends to diverging latent dimension and latent-class mixtures.","marker":"Hall et al. [2011a]"},{"why":"supplies the statistical EM contraction analysis used as the template for the linear-contraction lemma of the hybrid EM-GVA algorithm.","marker":"Balakrishnan et al. [2017]"},{"why":"supplies the multi-source EHR concept embeddings and feature search used to build V and the MS disability phenotyping input data.","marker":"Xiong et al. [2023]"}],"fun_headline_variants":["SCORE: Phenotyping EHRs with more unlabeled than labeled data","Label scarcity? Not for SCORE's EHR phenotyping","SCORE's error bound skips labeled-sample size","Unlabeled EHR data slash labeled data needs for phenotyping","SCORE turns unlabeled EHR data into phenotyping accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pre-trained code embeddings already span the true latent disease structure of the target cohort; if those external embeddings point in the wrong directions, the model is misspecified and the labeled-sample-free error rate no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["SCORE: Phenotyping EHRs with more unlabeled than labeled data","Label scarcity? Not for SCORE's EHR phenotyping","SCORE's error bound skips labeled-sample size","Unlabeled EHR data slash labeled data needs for phenotyping","SCORE turns unlabeled EHR data into phenotyping accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00114,"raw_usage":{"total_tokens":4811,"prompt_tokens":1100,"completion_tokens":3711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":716,"completion_tokens_details":{"reasoning_tokens":3628}},"tokens_in":716,"tokens_out":3711,"duration_ms":26613,"temperature":1.0,"reasoning_tokens":3628,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:47:34.805524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the PALM simulation with a known $V$, then rerun SCORE with $V$ altered by swapping one column for a random vector while holding $n$ fixed and doubling $N$; if the estimation error stops falling and stays above the predicted $\\psi q^{1/2}/N^{1/2}$ term, the span assumption is what carries the result.","supporting_citations":[{"cited_title":"The multivariate poisson-log normal distribution","cited_arxiv_id":null,"evidence_quote":"defines the multivariate Poisson-lognormal model that PALM extends to a latent-class mixture with covariates and known embeddings."},{"cited_title":"Variational inference for probabilistic poisson pca","cited_arxiv_id":null,"evidence_quote":"supplies the benchmark variational Poisson PCA method and the mean-field ELBO construction that SCORE adapts and extends to semi-supervised clustering."},{"cited_title":"A note on exploratory item factor analysis by singular value decomposition","cited_arxiv_id":null,"evidence_quote":"provides the two-step SVD baseline for item factor analysis that SCORE compares against and improves by modeling random effects and covariates."},{"cited_title":"Joint maximum likelihood estimation for high-dimensional exploratory item factor analysis","cited_arxiv_id":null,"evidence_quote":"gives the fixed-q joint maximum likelihood framework whose fixed-rank regime SCORE generalizes to diverging q."},{"cited_title":"Statistical guarantees for the em algorithm: From population to sample-based analysis","cited_arxiv_id":null,"evidence_quote":"supplies the statistical EM contraction analysis used as the template for the linear-contraction lemma of the hybrid EM-GVA algorithm."},{"cited_title":"Knowledge-driven online multimodal automated phenotyping system","cited_arxiv_id":null,"evidence_quote":"supplies the multi-source EHR concept embeddings and feature search used to build V and the MS disability phenotyping input data."}],"review_version":1}