Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Self-Supervised Representation Learning for Longitudinal Polypharmacy Patterns

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LOPEL, a two-stage self-supervised framework with an ATC-informed Gaussian process and a Wasserstein trajectory kernel, embeds longitudinal polypharmacy records into low-dimensional subject representations that recover latent clinical…

desk verdict LOPEL is a credible two-stage self-supervised embedding method for longitudinal medication data; the math is sound and the simulations are mostly convincing, but the uncertainty-based separation in Scenario 4 and the lack of replication/code keep the central claim from being fully established. read the letter →

arxiv 2608.06743 v1 pith:JXIBKIMP submitted 2026-08-07 stat.ME

Yunshan Duan , Eero Korpela , Yanxun Xu This is my paper · ORCID
classification stat.ME MSC 62H3062P10
keywords representationlearningself-supervisedlongitudinaldatapolypharmacyGaussianprocessWassersteindistanceATCclassificationclustering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

LOPEL is a self-supervised method for turning messy longitudinal medication records into compact patient-level summaries. Existing tools are either task-specific or cannot handle the combination of high-dimensional sparsity, irregular visit times, and the pharmacological structure among drugs. LOPEL first learns visit-level embeddings with a Gaussian process whose kernel encodes drug similarity through the ATC hierarchy, then learns subject-level embeddings from the trajectories of those visit embeddings using a Wasserstein-based similarity that includes estimation uncertainty. Simulation studies show it recovers known cluster structure where baseline methods fail, and an application to aging people with HIV yields four subgroups that differ in the timing and composition of polypharmacy. If correct, the framework gives clinicians and epidemiologists a reusable representation for risk stratification and comorbidity research.

What carries the argument

The load-bearing object is the trajectory kernel defined on each participant's Gaussian-process posterior: Stage 1 maps each visit's regimen to an embedding via a Gaussian process with an ATC-informed tree kernel (so that drugs sharing therapeutic ancestry are treated as similar), and Stage 2 fits a multitask GP per participant and evaluates the posterior on a common grid, giving a Gaussian $q_i = \mathcal{N}(m_i, \mathrm{diag}(\sigma_i^2))$. Similarity between participants is $K_W(q_i,q_j) = \exp\left(-W_2^2(q_i,q_j)/(2\tau^2)\right)$, where the diagonal Gaussian 2-Wasserstein distance is $W_2^2(q_i,q_j) = \|m_i - m_j\|^2 + \|\sigma_i - \sigma_j\|^2$. The mean term captures temporal dynamics while the variance term injects estimation uncertainty into the comparison, and exponentiating the negative squared distance yields a positive-definite covariance for the second-stage GPSSL embedding. The diagonal approximation is what makes the Wasserstein kernel valid and computable, at the cost of discarding cross-temporal and cross-dimension correlations.

What would settle it

Re-run LOPEL on the PWH cohort after stratifying or matching participants on the number of visits and see whether the four subgroups and their clinical profiles persist; if they dissolve or realign with visit-count strata, the uncertainty channel in the Wasserstein kernel is the operative clustering signal, not the medication trajectory.

Watch

Extended reading notes

Core claim

The central claim is that low-dimensional subject-level embeddings of longitudinal medication data can be learned without task-specific labels, and that these embeddings preserve the latent structure that drives medication trajectories. LOPEL represents each patient's medication history as a posterior distribution over a shared time grid and defines patient similarity through a diagonal-approximated 2-Wasserstein distance between those distributions, so that both the mean trajectory and its uncertainty enter the comparison. In simulations with sparse, irregularly sampled binary medication indicators, this representation recovers the true cluster structure with higher Adjusted Rand Index than dynamic time warping and hierarchical trajectory clustering baselines. In the application to 427 adults with HIV, clustering the learned embeddings identifies four groups that differ not only in total medication burden but in the age-specific accumulation of pain, hypertension, depression, airway, and cardiometabolic medications, which the authors interpret as clinically distinct longitudinal polypharmacy phenotypes.

Load-bearing premise

The kernel treats posterior uncertainty, which is mostly determined by how often a patient is observed, as meaningful signal; if visit frequency is correlated with disease severity or healthcare access, the learned subgroups may reflect follow-up intensity rather than medication-trajectory shape.

Editorial extensions

If this is right

  • Downstream tasks such as clustering, outcome prediction, and forecasting of future medication use can reuse the same subject-level embeddings instead of rebuilding task-specific models.
  • The four PWH subgroups give concrete risk-stratification targets: high-burden multisystem patients may need coordinated medication review, while younger low-burden patients can be monitored for trajectory shifts.
  • Because uncertainty enters the similarity metric, patients with sparse follow-up are not treated as having the same trajectory confidence as those with dense follow-up, which matters in real cohort data.
  • The same embedding pipeline could be applied to other high-dimensional, sparsely observed longitudinal measurements that carry a known hierarchical structure, such as laboratory panels or symptom inventories.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial: The uncertainty channel will reward or penalize participants based on follow-up intensity; in cohorts where visit frequency tracks disease severity, subgroups may partly encode care engagement rather than medication pattern, so a visit-count-adjusted analysis is a natural robustness check.
  • Editorial: The diagonal approximation discards correlations across time and medication domains; a low-rank covariance extension of the Wasserstein kernel could recover coordinated shifts in multiple conditions while preserving a positive-definite kernel.
  • Editorial: The same two-stage recipe could transfer to other sparse longitudinal signals with a known hierarchy, such as symptom checklists or diagnostic codes, if an appropriate tree kernel is supplied.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes LOPEL, a two-stage self-supervised framework for learning subject-level embeddings of longitudinal medication data. Stage 1 maps visit-level medication regimens to Euclidean embeddings using a Gaussian process prior with a positive-definite ATC tree kernel under a GPSSL variance/covariance loss; Stage 2 models each participant's trajectory with a multitask Gaussian process, represents the subject by the posterior distribution evaluated on a common age grid, and embeds these distributions using GPSSL with an exponentiated diagonal-2-Wasserstein kernel. The authors evaluate cluster recovery in four simulation scenarios against three baselines and apply the embeddings to identify four subgroups in a PWH cohort. The central claims are that LOPEL accurately recovers latent structure under realistic conditions and identifies clinically distinct longitudinal polypharmacy trajectories in PWH.

Significance. If the central empirical claims are sustained, LOPEL would be a meaningful contribution: it addresses a real gap in task-agnostic representations for sparse, irregular, high-dimensional medication data, and the methodological construction is internally sound. The ATC tree kernel is a valid positive-definite kernel, the diagonal-Wasserstein kernel reduces to a Euclidean RBF and is therefore positive definite, and the GPSSL loss is well-defined and genuinely self-supervised in that no outcome labels are used. The two-stage decomposition is principled, and the paper is appropriately careful about kernel validity. The main weaknesses are empirical: the simulation evidence rests on single replicates, and the 'uncertainty as signal' mechanism in Scenario 4 conflates trajectory-shape differences with differential follow-up density, which is not disentangled in the real-data application.

major comments (4)
  1. [Section 4.2, Table 1] The simulation studies report exactly one ARI value per scenario and method, generated from a single synthetic dataset. The central empirical claim that LOPEL 'accurately recovers latent structure' and outperforms comparators, especially in Scenario 4 (ARI 0.63 vs. best comparator 0.60), is therefore not supported by any estimate of sampling variability; a 0.03 difference from a single draw is within plausible Monte Carlo noise. I request multiple independent replicates (at least several seeds) with a report of the distribution of ARIs, and ideally a bootstrap or confidence interval, before the superiority claim can be considered established.
  2. [Section 3.3, Eq. (S.4), and Scenario 4 (Supp Table S2)] The subject-level trajectory kernel includes the term ||sigma_i - sigma_j||^2 in W2^2(qi, qj), where sigma_i are diagonal posterior standard deviations on a common age grid. For ages outside a participant's observed follow-up window the posterior variance reverts toward the GP prior variance, so this term is largely determined by visit count and follow-up length. Scenario 4 deliberately changes only the follow-up density for Cluster 3 (5-10 vs. 15-20 visits) while making its trajectory shapes qualitatively similar to Cluster 2 (Supp Table S2). Thus the claimed recovery in Scenario 4 is, by construction, an evaluation of the sigma term's ability to encode follow-up intensity rather than an evaluation of trajectory-shape recovery. In the PWH application, visit frequency is plausibly confounded with disease severity, healthcare access, and engagement; Section 5 neither reports visit counts or follow-up durations across PWH-C1-C4 nor tests whether the subgroups persist when the sigma term is removed from the kernel. Without such a sensitivity analysis, the interpretation of the four subgroups as 'clinically distinct longitudinal polypharmacy trajectories' is not established.
  3. [Section 3.4 and Supplementary Figures S2-S3] Embedding dimensions and loss weights are selected in each scenario by maximizing centered kernel alignment on the same simulated dataset that is then used to compute the reported ARI. Because the evaluation data are used for model selection, the reported ARI values are likely optimistically biased; the degree of overfitting cannot be assessed from the reported results. An independent validation split or a nested selection procedure should be used to evaluate out-of-sample recovery.
  4. [Section 5] The real-data analysis compares LOPEL only with KmeansDTW on total medication counts and on raw medication indicators. These baselines ignore both ATC structure and temporal dynamics, so they do not isolate the contribution of the Wasserstein uncertainty term. A more informative control would cluster participants on visit frequency or follow-up duration alone, or run LOPEL with the sigma term omitted (i.e., using only posterior means in the trajectory kernel), and then compare cluster stability. As reported, the application cannot distinguish 'clinically meaningful trajectories' from 'subgroups defined primarily by follow-up intensity.'
minor comments (5)
  1. [Section 4.1] There is a typo in the hyperparameter description: 'with gamma=10 and and a learning rate schedule' contains a duplicated 'and'.
  2. [Section 2] The overall dataset has a mean of 2.56 visits over an average follow-up of 3.30 years; it would be helpful to report the median and IQR of visit counts and follow-up duration, since Stage 2 posterior uncertainty is directly driven by these quantities.
  3. [Table S3] The selected embedding dimensions and loss weights vary substantially across scenarios (e.g., D=1 for Scenario 2 but D=2 for Scenario 4; d=5 for Scenario 2 but d=2 for Scenario 4). The authors should comment on whether this hyperparameter instability affects the comparability of ARI values across scenarios.
  4. [Section 5] After restricting to participants with at least three visits, the analytic sample has 427 PWH and 2,224 visits; reporting the mean/median number of visits per cluster would help calibrate how much posterior uncertainty remains in the real-data analysis.
  5. [Section 3.3] The text states that the posterior mean and covariance 'define a distribution over trajectories' but the subsequent comparison is performed on a fixed grid; the wording could be clarified to avoid implying a functional Wasserstein distance.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: LOPEL's representations are learned self-supervised and validated against independent simulation labels; only minor same-author method citation.

full rationale

The paper's derivation chain is self-contained. Stage 1 learns visit-level embeddings from a GP prior with an ATC-informed tree kernel and a self-supervised loss that only penalizes variance collapse and covariance redundancy; no outcome labels enter the loss. Stage 2 builds participant-specific multitask GPs, evaluates posterior means and marginal variances on a common grid, defines a diagonal-Gaussian Wasserstein trajectory kernel, and applies the same self-supervised GPSSL objective to obtain subject-level embeddings. In simulation, the reported ARI values compare the resulting clusters against independently generated true cluster labels, so the central claim 'LOPEL accurately recovers latent structure' is tested against external ground truth rather than being equivalent to the method's own inputs. The uncertainty (sigma) term in the Wasserstein kernel is a deliberate modeling choice: it uses posterior variance, which depends on observation density, and Scenario 4 explicitly includes differential follow-up as part of the cluster-generating mechanism. That is a construct-validity question (does uncertainty reflect clinical trajectory shape or follow-up intensity in real data?), not a circularity, because the paper does not define the target clusters as the output of its own kernel. The only self-referential element is the citation to the first author's prior GPSSL work [16] for the generalized Bayesian posterior and sparse variational inference; this is a method citation and the LOPEL-specific kernels and trajectory construction are evaluated on their own, so it is not load-bearing. Hyperparameter selection via centered kernel alignment against the target kernel is a model-selection device rather than a fitted prediction. Overall, no step reduces by construction to its inputs, and the minor self-citation does not undermine the independent content of the contribution.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several domain assumptions rather than on new physical postulates. The most important is the clinical validity of the ATC hierarchy and the RxRisk mapping, which shape both the kernel and the interpretation of clusters. The method also imports GPSSL as a trusted framework. Several hyperparameters are selected by alignment to the model's own kernels, and the final values are not all reported, which complicates independent reproduction. No invented entities are introduced.

free parameters (7)
  • ATC level weights w_l = (0.05, 0.30, 0.30, 0.15, 0.10)
    Hand-chosen in simulation to emphasize therapeutic and pharmacological subgroup levels; not fitted to labels but affects the similarity geometry.
  • Stage 1 dimension D and loss weights (cV, cC) = Scenario-dependent, e.g., D=2, (50,10) in Scenario 1; see Table S3
    Selected by centered kernel alignment against the ATC kernel itself; varies by scenario.
  • Stage 2 dimension d and loss weights (cV, cC) = Scenario-dependent, e.g., d=3, (10,10) in Scenario 1; real data d=3, (50,10)
    Selected by centered kernel alignment against the trajectory kernel; varies by scenario.
  • gamma (variance target in GPSSL loss) = not reported
    Controls the variance-promoting term in the Stage 1 and Stage 2 losses; value is inherited from GPSSL [16] but not stated in this paper.
  • tau (lengthscale of Wasserstein trajectory kernel) = learned, not reported
    Sets the scale of the exponential trajectory kernel; optimized during Stage 2 but final values not given.
  • GP kernel hyperparameters (output scale, RBF lengthscales, coregionalization B, noise Sigma_eps) = estimated, not reported
    Learned via marginal likelihood in Stage 2 and via ELBO in Stage 1; values not reported, affecting reproducibility.
  • RBF lengthscale gamma in Stage 2 temporal kernel = 10 (simulation)
    Set to 10 for subject-level embeddings in simulation; real-data value not reported.
assumptions (6)
  • domain assumption The ATC hierarchy encodes clinically meaningful pharmacologic similarity, so shared ATC ancestry is a valid similarity for medication regimens.
    Central to Stage 1 tree kernel (Section 3.2); if false, the kernel geometry misleads the embeddings.
  • domain assumption The generalized Bayesian posterior with GPSSL loss is a valid representation learning objective.
    Adopted from Duan and Williamson [16], Section 3.1; the paper does not re-derive or independently validate this framework.
  • domain assumption RxRisk-based medication-to-condition mapping is valid for interpreting cluster differences.
    Used in Section 2 and for interpreting PWH clusters; misclassification could create apparent comorbidity differences (acknowledged in Discussion).
  • domain assumption Diagonal approximation to posterior covariance retains sufficient signal for trajectory similarity.
    Section 3.3; the approximation discards temporal and cross-dimension correlations, which could matter when uncertainty is driven by sparse observations.
  • domain assumption Individual multitask GP fits with pooled hyperparameters are identifiable from sparse visits.
    Stage 2 models each participant's trajectory with an ICM GP; with as few as 3 visits, posterior distributions may be dominated by priors.
  • domain assumption Simulation data generated from condition-level prevalence curves mimic realistic medication use.
    Section 4.1; medications are assigned to conditions and generated from Bernoulli draws, which may not capture prescribing dynamics, drug-drug interactions, or discontinuation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Representation Learning for Longitudinal Polypharmacy Patterns." pith.science (2026). https://pith.science/paper/JXIBKIMP

@misc{pith2026260806743,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Representation Learning for Longitudinal Polypharmacy Patterns},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JXIBKIMP}},
  note         = {Machine review of arXiv:2608.06743}
}
read the original abstract

Polypharmacy, commonly defined as the concurrent use of multiple medications, is increasingly prevalent in aging populations and is associated with adverse health outcomes. Motivated by longitudinal studies of aging people with HIV (PWH), we study how medication use evolves over time and characterizes multimorbidity patterns. Longitudinal medication data present substantial methodological challenges, including high dimensionality, sparsity, irregular observation times, and structured pharmacologic relationships among medications. Existing approaches are typically task-specific and lack a unified framework for learning general-purpose representations of medication trajectories. We propose LOPEL (LOngitudinal Polypharmacy Embedding Learning), a two-stage self-supervised framework for learning low-dimensional representations of longitudinal medication data. In the first stage, visit-level embeddings are learned using a Gaussian process model that incorporates pharmacologic similarity through the Anatomical Therapeutic Chemical hierarchy. In the second stage, subject-level embeddings are constructed by modeling trajectories over time and defining similarity through a Wasserstein-based representation capturing temporal dynamics and uncertainty. Simulation studies demonstrate that LOPEL accurately recovers latent structure under realistic conditions with high-dimensional sparsity and irregular sampling. In an application to aging cohorts of PWH, LOPEL identifies clinically meaningful subgroups that differ in the timing, composition, and progression of medication use, highlighting heterogeneity relevant for risk stratification and clinical management.

Figures

Figures reproduced from arXiv: 2608.06743 by the authors.

Figure 1
Figure 1. Flowchart of the two stage construction of LOPEL. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the ATC hierarchy and the proposed tree-feature kernel [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. True cluster-specific condition prevalence curves [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Simulation results under Scenario 1 using LOPEL. (a) Gaussian process regression of [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Longitudinal medication patterns for PWH subgroups identified by LOPEL. (a) [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    Polypharmacy, hospitalization, and mortality risk: a nationwide cohort study.Scientific Reports, 10(1): 18964, 2020

    Tae Ik Chang, Haeyong Park, Dong Wook Kim, Eun Kyung Jeon, Connie M Rhee, Kamyar Kalantar-Zadeh, Ea Wha Kang, Shin-Wook Kang, and Seung Hyeok Han. Polypharmacy, hospitalization, and mortality risk: a nationwide cohort study.Scientific Reports, 10(1): 18964, 2020

    work page 2020

  2. [2]

    Prevalence and factors associ- ated with polypharmacy: a systematic review and meta-analysis.BMC Geriatrics, 22(1): 601, 2022

    Mahin Delara, Lauren Murray, Behnaz Jafari, Anees Bahji, Zahra Goodarzi, Julia Kirkham, Mohammad Chowdhury, and Dallas P Seitz. Prevalence and factors associ- ated with polypharmacy: a systematic review and meta-analysis.BMC Geriatrics, 22(1): 601, 2022. 23

    work page 2022

  3. [3]

    Meredith Greene, Michael A Steinman, Ian R McNicholl, and Victor Valcour. Polyphar- macy, drug–drug interactions, and potentially inappropriate medications in older adults with human immunodeficiency virus infection.Journal of the American Geriatrics Society, 62(3):447–453, 2014

    work page 2014

  4. [4]

    Comparison of overall and comorbidity-free life expectancy between insured adults with and without HIV infection, 2000-2016.JAMA Network Open, 3(6):e207954, 2020

    Julia L Marcus, Wendy A Leyden, Stacey E Alexeeff, Alexandra N Anderson, Rulin C Hechter, Haihong Hu, Jennifer O Lam, William J Towner, Qing Yuan, Michael A Horberg, et al. Comparison of overall and comorbidity-free life expectancy between insured adults with and without HIV infection, 2000-2016.JAMA Network Open, 3(6):e207954, 2020

    work page 2000

  5. [5]

    Non- AIDS-defining comorbidities impact health related quality of life among older adults living with HIV.Frontiers in Medicine, 11:1380731, 2024

    Alice Zhabokritsky, Marina Klein, Mona Loutfy, Giovanni Guaraldi, Nisha Andany, Silvia Guillemi, Julian Falutz, Gordon Arbess, Darrell HS Tan, and Sharon Walmsley. Non- AIDS-defining comorbidities impact health related quality of life among older adults living with HIV.Frontiers in Medicine, 11:1380731, 2024

    work page 2024

  6. [6]

    Degree of polypharmacy and cognitive function in older women with HIV.AIDS Research and Human Retroviruses, 38(7):571–579, 2022

    Leah H Rubin, Ava G Neijna, Qiuhu Shi, Donald R Hoover, Bani Tamraz, Kathryn Anastos, Andrew Edmonds, Margaret A Fischl, Deborah Gustafson, Pauline M Maki, et al. Degree of polypharmacy and cognitive function in older women with HIV.AIDS Research and Human Retroviruses, 38(7):571–579, 2022

    work page 2022

  7. [7]

    Anatomical therapeutic chemical classification system (ATC)

    Gerhard Nahler. Anatomical therapeutic chemical classification system (ATC). InDic- tionary of pharmaceutical medicine, pages 8–8. Springer, 2009

    work page 2009

  8. [8]

    Group- based multi-trajectory modeling.Statistical Methods in Medical Research, 27(7):2015– 2023, 2018

    Daniel S Nagin, Bobby L Jones, Valeria Lima Passos, and Richard E Tremblay. Group- based multi-trajectory modeling.Statistical Methods in Medical Research, 27(7):2015– 2023, 2018

    work page 2015

Show all 34 references
  1. [9]

    A latent class mixed model for analysing biomarker trajectories with irregularly scheduled observations.Statistics in Medicine, 19(10):1303–1318, 2000

    Haiqun Lin, Charles E McCulloch, Bruce W Turnbull, Elizabeth H Slate, and Larry C Clark. A latent class mixed model for analysing biomarker trajectories with irregularly scheduled observations.Statistics in Medicine, 19(10):1303–1318, 2000

    2000

  2. [10]

    Dynamic programming algorithm optimization for spoken word recognition.IEEE Transactions on Acoustics, Speech, and Signal Processing, 26(1): 43–49, 2003

    Hiroaki Sakoe and Seibi Chiba. Dynamic programming algorithm optimization for spoken word recognition.IEEE Transactions on Acoustics, Speech, and Signal Processing, 26(1): 43–49, 2003

    2003

  3. [11]

    Bayesian analysis for mixture of latent variable hidden Markov models with multivariate longitudinal data.Computational Statistics & Data Analysis, 132:190–211, 2019

    Ye-Mao Xia and Nian-Sheng Tang. Bayesian analysis for mixture of latent variable hidden Markov models with multivariate longitudinal data.Computational Statistics & Data Analysis, 132:190–211, 2019

    2019

  4. [12]

    Model-based clustering of high-dimensional longitudi- nal data via regularization.Biometrics, 79(2):761–774, 2023

    Luoying Yang and Tong Tong Wu. Model-based clustering of high-dimensional longitudi- nal data via regularization.Biometrics, 79(2):761–774, 2023. 24

    2023

  5. [13]

    A sparse factor model for clustering high- dimensional longitudinal data.Statistics in Medicine, 43(19):3633–3648, 2024

    Zihang Lu and Noirrit Kiran Chandra. A sparse factor model for clustering high- dimensional longitudinal data.Statistics in Medicine, 43(19):3633–3648, 2024

    2024

  6. [14]

    Longitudinal self-supervised learning.Medical Image Analysis, 71:102051, 2021

    Qingyu Zhao, Zixuan Liu, Ehsan Adeli, and Kilian M Pohl. Longitudinal self-supervised learning.Medical Image Analysis, 71:102051, 2021

    2021

  7. [15]

    Deep representation learning for clustering longitudinal survival data from electronic health records.Nature Communications, 16(1):2534, 2025

    Jiajun Qiu, Yao Hu, Li Li, Abdullah Mesut Erzurumluoglu, Ingrid Braenne, Charles Whitehurst, Jochen Schmitz, Jatin Arora, Boris Alexander Bartholdy, Shrey Gandhi, et al. Deep representation learning for clustering longitudinal survival data from electronic health records.Natur...

    2025

  8. [16]

    Self-Supervised Learning with Gaussian Processes

    Yunshan Duan and Sinead Williamson. Self-Supervised Learning with Gaussian Processes. arXiv preprint arXiv:2512.09322, 2025

    2025

  9. [17]

    R. K. Heaton, D. B. Clifford, D. R. Franklin, S. P. Woods, C. Ake, F. Vaida, R. J. Ellis, S. L. Letendre, T. D. Marcotte, J. H. Atkinson, M. Rivera-Mindt, O. R. Vigil, M. J. Taylor, A. C. Collier, C. M. Marra, B. B. Gelman, J. C. McArthur, S. Morgello, D. M. Simpson, J. A. McC...

    2010

  10. [18]

    The National NeuroAIDS Tissue Consortium: a new paradigm in brain banking with an emphasis on infectious disease.Neuropathology and Applied Neurobiology, 27(4):326–335, 2001

    S Morgello, BB Gelman, PB Kozlowski, HV Vinters, E Masliah, M Cornford, W Cavert, C Marra, I Grant, and EJ Singer. The National NeuroAIDS Tissue Consortium: a new paradigm in brain banking with an emphasis on infectious disease.Neuropathology and Applied Neurobiology, 27(4):32...

    2001

  11. [19]

    Welcome to the HNRP.https://hnrp

    HIV Neurobehavioral Research Program. Welcome to the HNRP.https://hnrp. hivresearch.ucsd.edu/index.php, 2026

    2026

  12. [20]

    Kapusnik-Uner, Thang Nguyen, and Olivier Bodenreider

    Lee Peters, Joan E. Kapusnik-Uner, Thang Nguyen, and Olivier Bodenreider. An Approx- imate Matching Method for Clinical Drug Names.AMIA Annual Symposium Proceedings, 2011:1117–1126, 2011. ISSN 1942-597X

    2011

  13. [21]

    RxNorm: prescription for electronic drug information exchange.IT Professional, 7(5):17–23, 2005

    Simon Liu, Wei Ma, Robin Moore, Vikraman Ganesan, and Stuart Nelson. RxNorm: prescription for electronic drug information exchange.IT Professional, 7(5):17–23, 2005

    2005

  14. [22]

    ATC.WHO Drug Information, 31(4):629–634, 2017

    World Health Organization et al. ATC.WHO Drug Information, 31(4):629–634, 2017

    2017

  15. [23]

    National Library of Medicine

    U.S. National Library of Medicine. Rxclass.https://rxnav.nlm.nih.gov/RxClass/, 2026. 25

    2026

  16. [24]

    Selective serotonin reuptake inhibitors

    Dee Lochmann and Tara Richardson. Selective serotonin reuptake inhibitors. InAntide- pressants: From biogenic amines to new mechanisms of action, pages 135–144. Springer, 2019

    2019

  17. [25]

    Pratt, Mhairi Kerr, John D

    Nicole L. Pratt, Mhairi Kerr, John D. Barratt, Anna Kemp-Casey, Lisa M. Kalisch Ellett, Emmae Ramsay, and Elizabeth Ellen Roughead. The validity of the Rx-Risk Comor- bidity Index using medicines mapped to the Anatomical Therapeutic Chemical (ATC) Classification System.BMJ Ope...

    2018 doi

  18. [26]

    Validity of the Updated Rx-Risk Index as a Disease Identification and Risk-Adjustment Tool for Use in Observational Health Studies.Clinical Interventions in Aging, 20:309–323, 2025

    Imaina Widagdo, Mhairi Kerr, Lisa Kalisch Ellett, Clement Schlegel, Elham Sadeqzadeh, Alvin Wang, Allison Louise Clarke, and Nicole Pratt. Validity of the Updated Rx-Risk Index as a Disease Identification and Risk-Adjustment Tool for Use in Observational Health Studies.Clinica...

    2025 doi

  19. [27]

    A general framework for updating belief distributions.Journal of the Royal Statistical Society Series B: Statistical Methodology, 78(5):1103–1130, 2016

    Pier Giovanni Bissiri, Chris C Holmes, and Stephen G Walker. A general framework for updating belief distributions.Journal of the Royal Statistical Society Series B: Statistical Methodology, 78(5):1103–1130, 2016

    2016

  20. [28]

    Generalized variational inference: Three arguments for deriving new posteriors.arXiv preprint arXiv:1904.02063, 2019

    Jeremias Knoblauch, Jack Jewson, and Theodoros Damoulas. Generalized variational inference: Three arguments for deriving new posteriors.arXiv preprint arXiv:1904.02063, 2019

    1904 arXiv

  21. [29]

    Scalable variational Gaussian process classification

    James Hensman, Alexander Matthews, and Zoubin Ghahramani. Scalable variational Gaussian process classification. InArtificial intelligence and statistics, pages 351–360. PMLR, 2015

    2015

  22. [30]

    Similarity of neural network representations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. InInternational conference on machine learning, pages 3519–3529. PMlR, 2019

    2019

  23. [31]

    ClusterMLD: An efficient hierarchical cluster- ing method for multivariate longitudinal data.Journal of Computational and Graphical Statistics, 32(3):1131–1144, 2023

    Junyi Zhou, Ying Zhang, and Wanzhu Tu. ClusterMLD: An efficient hierarchical cluster- ing method for multivariate longitudinal data.Journal of Computational and Graphical Statistics, 32(3):1131–1144, 2023

    2023

  24. [32]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis.Journal of Computational and Applied Mathematics, 20:53–65, 1987

    Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis.Journal of Computational and Applied Mathematics, 20:53–65, 1987

    1987

  25. [33]

    Ann N Do, Eli S Rosenberg, Patrick S Sullivan, Linda Beer, Tara W Strine, Jeffrey D Schulden, Jennifer L Fagan, Mark S Freedman, and Jacek Skarbinski. Excess burden of depression among HIV-infected persons receiving medical care in the United States: data 26 from the medical m...

    2014

  26. [34]

    Self-Supervised Representation Learning for Longitudinal Polypharmacy Patterns

    Kristina Crothers, Adeel A Butt, Cynthia L Gibert, Maria C Rodriguez-Barradas, Stephen Crystal, Amy C Justice, Veterans Aging Cohort 5 Project Team, et al. Increased COPD among HIV-positive compared to HIV-negative veterans.Chest, 130(5):1326–1333, 2006. 27 Supplementary Mater...

    2006

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.