{"id":"e5be800f-84d6-46e0-bf64-8f912f7b436a","arxiv_id":"2508.20327","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A latent factor point process model plus Fourier spectral embeddings gives new patient-level representations for EHR classification and clustering, but the stated theoretical guarantees contain a diverging error term.","lead":"This paper proposes a statistical model in which a patient's medical codes are driven by a few hidden disease processes, and uses the frequency content of code co-occurrence to build a low-dimensional patient profile. The authors prove bounds for classification and clustering on these profiles, though a key bound contains an error that grows with observation time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central eigenvalue-separation identity is false: for non-orthogonal W, the top eigenvalues of W D_i W^+ are not a fixed linear function of μ_i, so Lemma 3.4 and Theorems 3.10/3.13 are unsupported.","rationale":"Read in good faith, the paper's contribution has two parts: an empirically plausible embedding with simulations and an EHR application, and a theoretical guarantee that the embedding separates latent classes. The theory is the advertised reason for preferring Fourier-Eigen embeddings over counts, PMI, and BERT, and Theorems 3.10 and 3.13 are explicit. The load-bearing step is Lemma 3.4, which converts assumed separation in μ(g) into separation in the eigenvalue embeddings. That step is reached through the 'H D_i' assertion, which is a mathematical identity that fails as soon as W has non-orthogonal columns. Because Assumption 3.3 explicitly permits non-orthogonal W and the text itself notes W^+W≠I, this is not a stray edge case. A concrete two-dimensional example shows no fixed linear map can exist; the eigenvalues depend on μ nonlinearly through D^{1/2} W^+W D^{1/2}. The simulations use k=2 and random coefficient matrices that are generically non-orthogonal, so the empirical success cannot be explained by the stated theory. The T^{6/5} term in Theorem 3.10 is an additional internal inconsistency: the displayed bound grows with T rather than shrinking, so even the final inequality as written does not deliver the advertised consistency; although this may be a typographical error, it reinforces that the theorem statements have not been carefully verified. The empirical method might survive a repaired proof or a weaker separation condition, but as submitted the central theoretical claim is unsupported. I agree with the reader's rejection and do not see a reason to change the verdict.","tokens_in":20778,"tokens_out":10040,"duration_ms":92018,"concrete_test":"Numerically test the asserted identity with k=2 and W chosen so W^+W=[[1,0.6],[0.6,1]]. Compute the top two eigenvalues f(μ) of W diag(μ) W^+ for μ=(1,0), (0,1), and (1,1). A fixed linear map H would require f(1,1)=f(1,0)+f(0,1)=(1,1), while the actual eigenvalue pair is (1.6,0.4). Confirming this mismatch directly refutes the H D_i identity and therefore the derivation of Lemma 3.4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After Eq. (2.10) the paper asserts that for full-column-rank W(ξ), the top k eigenvalues of W(ξ) D_i W(ξ)^+ can be written as H D_i for a fixed invertible H. This is false unless W^+W is diagonal, and Assumption 3.3 only requires full column rank. Nonzero eigenvalues of W D_i W^+ coincide with those of D_i^{1/2} G D_i^{1/2}, G=W^+W, which is a nonlinear function of μ_i. Example: k=2, G=[[1,c],[c,1]] with c≠0. For μ=(1,0) the sorted eigenvalues are (1,0); for μ=(0,1) they are (0,1); for μ=(1,1) they are (1+c,1-c). Any fixed H would have columns (1,0) and (0,1), forcing H(1,1)=(1,1), a contradiction. Lemma 3.4's lower bound in (3.2) is the bridge translating separation in μ(g) into separation of the embeddings, and Theorems 3.10 and 3.13 depend on that lemma through conditions (3.9) and (3.12). Since Assumption 3.3 explicitly permits non-orthogonal W, the advertised guarantees are not established for the matrices allowed by the paper's own assumptions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a latent factor point process model for high-dimensional EHR event streams, in which an observed code process has conditional intensity driven by an unobserved low-dimensional homogeneous Poisson process. The authors define Fourier-Eigen embeddings as the top k eigenvalues of the Fourier-transformed cross-covariance matrix of the observed process, and claim theoretical guarantees for downstream classification by empirical risk minimization and clustering by K-means with spectral initialization. The theoretical results are supplemented by simulation studies and an application to clustering Alzheimer's disease patients at UPMC, with nursing home admission as a progression outcome.","tokens_in":21143,"tokens_out":4776,"duration_ms":48208,"significance":"If the theoretical claims were correct, the paper would contribute a principled spectral method for patient-level temporal EHR representations, with explicit generalization and clustering guarantees. The modeling idea is appealing, the algorithmic pipeline is clearly stated, and the empirical section is extensive, including a real-data analysis with interpretable cluster characterizations. However, the central theoretical bridge from latent-class separation to embedding separation rests on an eigenvalue linearity claim that is false under the paper's own assumptions, and the main risk bound in Theorem 3.10 grows with observation time rather than vanishing. These are load-bearing issues for the advertised guarantees, so the paper's core theoretical contribution is not currently established.","major_comments":[{"comment":"The claim after Eq. (2.10) that, for full-column-rank W(ξ), the top k eigenvalues of W(ξ) D_i W(ξ)^† can be written as H D_i for a fixed invertible H is false for non-orthogonal W. The nonzero eigenvalues of W D W^† coincide with those of D^{1/2} G D^{1/2} with G = W^† W, which is a nonlinear function of D unless G is diagonal. For k=2 and G = [[1,c],[c,1]] with c≠0, D=diag(1,0) gives sorted eigenvalues (1,0), D=diag(0,1) gives (0,1), and D=diag(1,1) gives (1+c,1-c); no fixed H can reproduce all three because any H mapping (1,0) and (0,1) to the first two cases is the identity. Since Assumption 3.3 permits arbitrary full-column-rank W, the separation between population embeddings asserted in Lemma 3.4, and hence the signal conditions (3.9) and (3.12) for Theorems 3.10 and 3.13, are unsupported for the matrices allowed by the paper's own assumptions.","section":"Eq. (2.10) and Lemma 3.4"},{"comment":"The second term on the right-hand side of Eq. (3.10), 2 c2 d^2 T^{6/5} exp(c3 T^{-1/5}) diam(G), grows without bound as T increases because the exponential tends to 1. Thus the claimed generalization bound does not improve with longer observation windows and is vacuous for large T. This is not a harmless presentation issue: it contradicts the paper's narrative that longer trajectories improve the embeddings, and it stands in contrast to Theorem 3.13, whose probability bound has the complementary factor exp(-c3 T^{1/5}_{min}) and tends to 1 as T_min grows. The scaling appears to be a sign or bandwidth error and must be corrected before the theorem can be interpreted as a valid risk bound.","section":"Theorem 3.10, Eq. (3.10)"},{"comment":"Lemma 3.4 is the sole bridge translating separation in the latent intensity vectors μ(g) into separation of the Fourier-Eigen embeddings, but its proof is not included in the manuscript; the text refers to a supplement that is not available in the submitted version. Given the false linearity premise identified above, the proof cannot be reconstructed from the stated assumptions, so the conditions (3.9) and (3.12) are unverified. This missing support is load-bearing because both main theorems depend on the lemma.","section":"Lemma 3.4 and Appendix"}],"minor_comments":[{"comment":"The factorization in Eq. (2.10) uses the transpose (W(ξ))^⊤, but for complex-valued Fourier transforms the Hermitian property stated in Lemma 3.2 requires the conjugate transpose W(ξ)^†. The notation should be corrected for consistency with the definition of A† in Section 1.2.","section":"Eq. (2.10)"},{"comment":"The theory assumes compactly supported transfer functions (Assumption 3.6), but the main simulation uses ω_jℓ(t) = a_jℓ e^{-t^2/2}, which is not compactly supported on R_+; only the robustness study uses compactly supported kernels. The primary simulations therefore do not operate in the setting covered by the theorems.","section":"Assumption 3.6 vs. Section 4"},{"comment":"The sentence defining M_i as a homogeneous Poisson process 'on R^k with intensity vector μ' conflates the time domain of the point process with the dimension k of its components; the process is k-dimensional over time [0,T_i]. This wording should be clarified.","section":"Section 2.1, Eq. (2.2)"},{"comment":"There are several minor typographical issues, including 'has beget' in the first sentence of the Introduction, a missing comma in the sentence before Eq. (2.7), and an apparent paragraph break after Eq. (2.9). These do not affect the technical content but should be cleaned up.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper has an appealing empirical package and a clearly motivated model, but the theoretical core is not sound as written. The eigenvalue linearity claim after Eq. (2.10) is false under Assumption 3.3, and Theorem 3.10's bound diverges with T. These are not localized presentation issues; they invalidate the main advertised guarantees and would require new assumptions or a different proof strategy to repair. I recommend rejection, though the authors could resubmit a substantially revised version with a corrected spectral separation argument and a valid risk bound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper has a plausible modeling idea and some decent empirical work, but its main theoretical claim is built on a false linearity assertion. That is a load-bearing flaw, not a fixable typo.\n\nWhat is genuinely new is the patient-level Fourier-Eigen embedding: using the leading eigenvalues of the Fourier-transformed cross-covariance matrix as a low-dimensional patient representation. That is a reasonable and under-explored direction for EHR data, and the authors are honest about the model being a special case of Tang and Li (2023). The simulation study and the Alzheimer's cohort analysis are competently done, and the empirical results suggest the method can work in practice. The code and data are not shipped, but the experiments are reproducible in spirit.\n\nThe problem starts after Eq. (2.10). The paper asserts that for full-column-rank W(ξ), the top k eigenvalues of W D_i W^† can be written as H D_i for a fixed invertible H. That is false unless W^† W is diagonal. For non-orthogonal W, the nonzero eigenvalues of W D_i W^† equal those of D_i^{1/2} G D_i^{1/2} with G = W^† W, which is nonlinear in D_i. A simple k=2 counterexample with G = [[1,c],[c,1]] shows no fixed H can exist when c≠0. Assumption 3.3 only requires full column rank, so the paper's own assumptions allow this failure. Lemma 3.4, which translates separation in μ(g) into separation of the embeddings, depends on this linearity, and Theorems 3.10 and 3.13 inherit the problem through conditions (3.9) and (3.12). The advertised generalization and exact-recovery guarantees are therefore not established.\n\nThere is a second, independent issue in Theorem 3.10. The bound contains 2 c2 d^2 T^{6/5} exp(c3 T^{-1/5}) diam(G), which grows without bound as T increases. A risk bound that worsens with more observation time is not a meaningful generalization guarantee, even if the signal condition also improves. This looks like a missing factor or a misassigned exponent, but as written it is another reason the theory does not support the conclusions.\n\nWho is this for? Methodologically interested biostatisticians and ML readers working on temporal EHR representations will find the embedding idea worth knowing, and the empirical comparison with PMI and BERT is useful. But the paper currently overclaims its theoretical support. I would send it to peer review, because the problem is important and the empirical core might survive a major revision. As is, it should not be accepted; the theory needs to be repaired or substantially weakened, and the risk bound needs to be fixed.\n\nIn short: the intuition is sound, the empirics are promising, but the math in Sections 3 does not hold up. A serious referee should engage with it, but the current version does not deliver what it promises.","headline":"Useful-looking EHR embedding under a latent factor point process model, but the central eigenvalue linearity claim is false, so the advertised classification and clustering guarantees don't hold as stated.","tokens_in":21644,"tokens_out":1440,"would_cite":false,"duration_ms":15597,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M15","62M30","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A few eigenvalues of the Fourier-transformed code co-occurrence matrix are a sufficient patient representation for classification and clustering under the paper's latent factor point process model.","keywords":["latent factor point process","Fourier-Eigen embedding","spectral density matrix","electronic health records","high-dimensional point process","patient representation","classification","clustering"],"falsifier":"Compute the eigenvalues of $W \\operatorname{diag}(\\mu) W^\\dagger$ for $W = [[1,0],[1,1]]$ at $\\mu = (1,1)$, $(1,2)$, $(2,1)$, and $(2,2)$: the eigenvalue shifts for the two unit directions do not combine linearly, e.g. the shift for $(1,1)$ is $(2.618, 0.382)$, not the sum of the shifts for $(1,0)$ and $(0,1)$, so no fixed $H$ can express the top eigenvalues as $H\\mu$. This is a concrete case where the linearity premise after Eq. (2.10) fails.","tokens_in":20581,"feed_emoji":"🩺","tokens_out":9182,"duration_ms":79235,"temperature":0.7,"pith_summary":"The paper tries to establish that patient-level temporal structure in electronic health records can be compressed into a small vector of spectral eigenvalues without losing the signal that separates clinical subgroups. It introduces a latent factor point process model in which thousands of observed medical codes are driven by a few unobserved Poisson processes, and defines the Fourier-Eigen embedding as the leading eigenvalues of the Fourier-transformed cross-covariance matrix of a patient's codes. Under this model, the paper proves that classifiers trained on these embeddings attain bounded risk and that spectral-initialized K-means recovers the true subgroup labels exactly when the latent intensity vectors are sufficiently separated. If correct, this gives EHR analysts a nonparametric, theoretically grounded alternative to count-based summaries and code-level embeddings, and simulations plus an Alzheimer's disease cohort analysis indicate the embeddings capture clinically meaningful heterogeneity.","feed_headline":"Eigenvalues of EHR code spectra separate patient subgroups","feed_subtitle":"A few Fourier-domain eigenvalues of code co-occurrence provably drive classification and clustering of patient trajectories.","key_machinery":"The load-bearing object is the spectral density matrix $F\\{V^{N_i}\\}(\\xi) = W(\\xi) D_i W(\\xi)^\\dagger$, where $W(\\xi)$ collects the Fourier transforms of the shared transfer functions $\\omega_{j\\ell}$ and $D_i = \\operatorname{diag}(\\mu_i)$ collects the latent Poisson intensities of patient $i$. Its top $k$ eigenvalues are the Fourier-Eigen embedding used for downstream tasks. The mechanism runs through three steps: Proposition 2.1 shows the cross-covariance inherits the low rank of the latent process; Lemma 3.4 translates differences between latent intensity vectors into $\\ell^2$ separation of the population embeddings; and Theorems 3.10 and 3.13 convert that separation into classifier risk bounds and exact clustering recovery. The kernel-smoothed cross-covariance estimator of Algorithm 4 supplies the finite-sample analogue of the population matrix.","core_discovery":"The central claim is that under the latent factor point process model (2.1)-(2.3), the cross-covariance matrix of the observed code process factors as $V^{N_i}(\\tau) = (\\omega \\star D_i \\star \\omega^\\top)(\\tau)$, with $D_i = \\operatorname{diag}(\\mu_i)$ the latent intensity matrix, so that its Fourier transform is $W(\\xi) D_i W(\\xi)^\\dagger$. The top $k$ eigenvalues of this spectral density matrix define the Fourier-Eigen embedding $f_i$, and the paper argues these eigenvalues inherit the subgroup differences in $\\mu^{(g)}$. Theorems 3.10 and 3.13 formalize the guarantee: an empirical risk minimizer on the estimated embeddings has expected loss bounded by an approximation term, a dimension-dependent estimation term of order $d^2 T^{6/5} \\exp(c T^{-1/5})$, and a Rademacher or VC complexity term, while spectral-initialized K-means recovers the latent class labels with high probability once the signal strength condition (3.12) holds.","pith_inferences":["The single-frequency choice $\\xi_0$ may be more consequential than the paper suggests: averaging embeddings across a band of frequencies would test whether the guarantees are robust to misspecification of the transfer functions and would likely stabilize the estimator.","A direct test of the latent Poisson-driver assumption would be to apply the embedding to EHR data with known self-exciting code patterns; if the cross-covariance factorization fails, the separation guarantees should degrade in a measurable way.","If the linearity premise behind Lemma 3.4 fails for non-orthogonal transfer matrices, a natural repair is to replace raw eigenvalues by singular values or to orthogonalize $W(\\xi)$; this would preserve the low-rank separation argument while changing the embedding's definition.","The Alzheimer's analysis uses nursing-home admission as a proxy for severity; validating the clusters against chart-reviewed severity labels would clarify whether the embedding discovers disease progression itself or merely correlates with care utilization."],"forward_implications":["Count-based EHR summaries are provably discarding temporal signal that the Fourier-Eigen embeddings retain, and the embeddings' classification performance improves with observation time $T$ while counts do not.","Only $k$ spectral eigenvalues are needed as features, so downstream classifiers can work in dimension $k$ instead of code dimension $d$, shrinking VC dimension and Rademacher complexity accordingly.","Unsupervised subgroup discovery is exact under a stronger signal condition than classification, so the gap between the two theorems quantifies the price of not observing labels.","In the Alzheimer's disease cohort, the two clusters formed from the embeddings have significantly different nursing-home admission trajectories, and the cluster membership carries a significant Cox hazard ratio, unlike the competing PMI and transformer-based embeddings."],"supporting_citations":[{"why":"Supplies the general cross-covariance factorization that Proposition 2.1 specializes to the latent factor point process model.","marker":"Tang and Li (2023)"},{"why":"Provides the kernel-smoothed cross-covariance estimator used in Algorithm 4 and the concentration bounds behind Theorems 3.10 and 3.13.","marker":"Chen et al. (2017)"},{"why":"Contributes the spectral-initialized K-means algorithm, Algorithm 3, whose exact recovery guarantee Theorem 3.13 relies on.","marker":"Kumar and Kannan (2010)"},{"why":"Defines the PMI embedding baseline and the log-linear latent variable model that the simulations and data analysis compare against.","marker":"Arora et al. (2016)"},{"why":"Supplies the convolution theorem that converts the lag-domain factorization into the spectral density matrix $W(\\xi) D_i W(\\xi)^\\dagger$.","marker":"Dym and McKean (1972)"},{"why":"Establishes the spectral analysis of point processes that motivates defining the embedding from the Fourier-transformed cross-covariance.","marker":"Bartlett (1963)"}],"fun_headline_variants":["EHR code spectra's top eigenvalues distinguish patient groups","A few eigenvalues of code co-occurrence split patients","Spectral density eigenvalues of EHR codes cluster patients","Low-rank code spectra yield patient subgroup separators","EHR code spectrum eigenvalues drive patient subtyping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that the top eigenvalues of the spectral matrix change linearly with the latent disease-process intensities for any full-rank transfer matrix, but eigenvalue maps are generally nonlinear unless the transfer columns are orthogonal, so separation in latent intensities may not survive in the embeddings.","fun_headline_variants_meta":{"raw":{"variants":["EHR code spectra's top eigenvalues distinguish patient groups","A few eigenvalues of code co-occurrence split patients","Spectral density eigenvalues of EHR codes cluster patients","Low-rank code spectra yield patient subgroup separators","EHR code spectrum eigenvalues drive patient subtyping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000435,"raw_usage":{"total_tokens":2207,"prompt_tokens":934,"completion_tokens":1273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1199}},"tokens_in":550,"tokens_out":1273,"duration_ms":9295,"temperature":1.0,"reasoning_tokens":1199,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:46:39.031131+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the eigenvalues of $W \\operatorname{diag}(\\mu) W^\\dagger$ for $W = [[1,0],[1,1]]$ at $\\mu = (1,1)$, $(1,2)$, $(2,1)$, and $(2,2)$: the eigenvalue shifts for the two unit directions do not combine linearly, e.g. the shift for $(1,1)$ is $(2.618, 0.382)$, not the sum of the shifts for $(1,0)$ and $(0,1)$, so no fixed $H$ can express the top eigenvalues as $H\\mu$. This is a concrete case where the linearity premise after Eq. (2.10) fails.","supporting_citations":[],"review_version":1}