Pith. sign in

REVIEW 4 major objections 6 minor 23 references

Proactive HIV Care: AI-Based Comorbidity Prediction from Routine EHR Data

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Demographic-aware models outperform demographic-unaware models across all six classifiers for HIV comorbidity prediction, and lab data alone already encodes gender and age.

desk verdict A routine feature-ablation study whose recoverability claim is contaminated because the 'lab-only' arm includes social/behavioral variables, and whose 'proactive' framing is untested. read the letter →

arxiv 2508.20133 v2 pith:SGJSRO5E submitted 2025-08-26 cs.CY

Solomon Russom , Dimitrios Kollias , Qianni Zhang This is my paper · ORCID
classification cs.CY
keywords HIVcomorbiditiesmulti-labelclassificationelectronichealthrecordsdemographic-awaremodelingdemographicrecoverabilityXGBoostclinicalAImacroF1
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

The paper tries to establish that explicitly including demographic information (age group, gender, race, continent of birth) alongside routine lab tests improves multi-label prediction of twelve ICD-10 comorbidity categories in people living with HIV. Across six classifiers, the demographic-aware versions consistently beat demographic-unaware versions on macro F1, with XGBoost best. The paper also shows that gender (92.8% F1) and age (45.4%) can be recovered from laboratory data alone, meaning the unaware models are not truly blind to demographics. A sympathetic reading: the performance gap is real, but the cleanest interpretation is that demographics add predictive context, not that the gap proves a specific mechanism.

What carries the argument

The carrying mechanism is the paired comparison of demographic-aware p(y | x, s) versus demographic-unaware p(y | x) models, where x is 30 lab features plus 3 social features and s is age group, gender, race, continent of birth. The recoverability experiments use XGBoost as a probe to measure how much of s can be reconstructed from lab data, quantifying proxy leakage. The central identity is that macro-F1 is averaged over the 12 ICD-10 blocks, so the comparison is a direct measure of whether explicit demographics add predictive information beyond what labs already encode.

What would settle it

Retrain the demographic-unaware models on a version that removes the three social features and compare macro F1 to the demographic-aware models; if the gap closes or reverses, the paper's attribution of the gain to explicit demographics would be falsified. A separate check: split the cohort temporally (train on 2012-2019, test on 2020-2023); if the demographic-aware advantage disappears under that external validation, the proactive claim fails.

Watch

Extended reading notes

Core claim

The central claim is that, for this cohort of about 2,200 HIV outpatients in South East London, feeding demographic and social attributes into standard ML/DL classifiers raises macro-F1 comorbidity prediction from 28.0-43.5% to 32.6-45.8% depending on model, with every model improving. The recoverability experiments show gender is nearly perfectly inferable from lab tests (92.8% F1) and age is substantially inferable (45.4%), while race and country of birth are weaker, which the paper reads as evidence that the predictive signal of demographics is already partly embedded in lab markers. The paper's conclusion is that responsible inclusion of demographics can improve automated comorbidity scr

Load-bearing premise

The comparison assumes the demographic-unaware arm is genuinely free of demographic information, but the social features (sexual orientation, commercial sex worker status, route of infection) are likely correlated with gender and age, so the unaware models are probably not truly unaware.

Editorial extensions

If this is right

  • If the result holds, demographic-aware models are preferable for clinical risk stratification when demographics are ethically and legally usable, improving detection especially for rare comorbidities like neoplasms and nervous-system disorders.
  • The near-perfect gender recoverability implies that any model trained on these lab markers will indirectly use gender-like information even when gender is excluded, so fairness audits must treat lab panels as potential proxies for protected attributes.
  • The consistent performance gains across all six models, including deep learning, suggests the benefit is not a quirk of one algorithm but a property of the data.
  • The Z00-Z99 category, present in 87% of patients, shows no gain from demographics, indicating the benefit concentrates in less prevalent, more clinically specific categories.

Reading between the lines

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

  • The comparison is not a clean test of demographics alone, because the unaware arm still includes sexual orientation, commercial sex worker status, and route of infection, which correlate with gender and age; retraining with those removed would establish the size of a pure demographic effect.
  • The study reports no temporal split, so the proactive label is not yet supported; a hold-out evaluation on later calendar years would test whether the models generalize forward in time the way a screening tool must.
  • If gender is 92.8% recoverable, the performance gap between aware and unaware models may partly reflect the models' difficulty in reconstructing gender from labs; giving the model gender directly removes a noisy inference step.
  • The ICD-10 top-level categories are broad; a more clinically actionable test would use specific diagnoses (e.g., diabetes, CKD stage, depression) rather than high-level blocks.
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 / 6 minor

Summary. The paper evaluates six machine/deep learning models (LR, RF, XGBoost, LightGBM, MLP, TabNet) for multi-label comorbidity prediction in a cohort of about 2,200 HIV-positive patients from South East London, using 30 laboratory markers plus demographic/social attributes. It compares demographic-aware models (lab + demographic/social inputs) with demographic-unaware models (lab + social inputs, excluding demographic attributes), reporting that aware models consistently outperform unaware counterparts, with XGBoost achieving the highest macro-F1. The paper also reports 'demographic recoverability' experiments using XGBoost to infer gender, age, race, and continent of birth from 'lab tests only' inputs, and interprets the aware-vs-unaware performance gap as a direct consequence of demographic signal embedded in lab data. The recoverability experiments, however, are contaminated by the inclusion of social/behavioral features (sexual orientation, commercial sex worker status, route of infection) in the 'lab tests only' configuration, and the unaware arm also includes these social proxy variables. These issues undermine the paper's central interpretative claims, though the empirical benchmark itself may remain valid after reanalysis.

Significance. If the central claims were established, the paper would make a useful contribution to clinical AI for HIV care: it uses a real EHR cohort, compares a broad set of models, reports per-comorbidity performance, and raises fairness considerations around demographic leakage. The recoverability finding, in particular, would be practically important for fairness audits. However, the current evidence does not support the headline claim that gender and age are accurately inferable from laboratory data alone, because the 'lab-only' configuration includes social/behavioral features that correlate strongly with demographics. Similarly, the unaware arm is not cleanly free of demographic information. These are not mere presentation issues: they directly affect the paper's main interpretation. The paper's strengths are the multi-label benchmark and the per-category analysis; these could be preserved if the recoverability and unaware-setup definitions are corrected and uncertainty estimates are provided.

major comments (4)
  1. [4.2, Table 5, Abstract] The recoverability experiments labeled 'Lab Tests Only' are defined as using 'all available laboratory and social measurements while excluding any demographic variables.' Section 3.3 lists the social features as Sexual Orientation, Commercial Sex Worker status, and Route of Infection. These variables are strong demographic proxies: for example, commercial sex worker status and route of infection correlate with gender, and sexual orientation is a demographic attribute in its own right. Therefore the reported high recoverability (gender macro-F1 92.8%, age 45.4%) cannot be attributed to laboratory data alone. The Abstract's statement that 'gender and age can be accurately inferred from laboratory data' and Section 4.2's conclusion that 'the majority of demographic signal is already embedded in lab test results' are unsupported. The recoverability experiments should be rerun with strictly l
  2. [3.1, 3.3, Tables 3–4] The demographic-unaware models are described in Section 3.1 as using 'only the laboratory and social features,' and Section 3.3 includes Sexual Orientation, Commercial Sex Worker status, and Route of Infection among the social features. These are not neutral controls: they carry substantial demographic signal (CSW status is heavily gender-skewed, and route of infection correlates with gender and age). Consequently, the aware-vs-unaware comparison is not a clean test of whether demographic features drive the performance gain. It tests whether adding age/gender/race/continent on top of lab+social features helps. The conclusion 'demographic-aware approach outperformed' remains supported by Table 3, but the interpretation that the gap reflects demographic information per se is weakened. The authors should either exclude social/behavioral proxies from the unaware arm or explicitly analyze the
  3. [Table 3, §4.1] The central claim that 'the demographic-aware approach outperformed the demographic-unaware approach across all models' is supported only by point estimates averaged across folds, with no standard deviations, confidence intervals, or significance tests reported. The deltas are modest (e.g., XGB 45.8 vs 43.5, LGBM 44.7 vs 43.0, RF 40.7 vs 38.7). Without uncertainty quantification, it is impossible for a reader to assess whether these differences are reliable or within fold-to-fold noise. Since the paper's main conclusion rests on this comparison, the authors should report fold-level distributions, standard errors, or paired significance tests across folds.
  4. [§3.3, §5, Abstract] The paper repeatedly frames the task as 'proactive' and 'early detection' of comorbidities (Abstract: 'early detection is often limited by symptom-driven screening'; Section 5: 'future research should explore richer multimodal and longitudinal data integration'). However, the experimental design uses the full EHR records from 2012–2023 without any temporal split: features and labels appear to be drawn from the same patient episodes. The models therefore predict the presence or absence of ICD-coded diagnoses contemporaneously, not future diagnoses. To support the 'proactive' claim, a temporal split (e.g., using baseline lab features to predict diagnoses occurring in a subsequent window) is required. This is a missing element of the evaluation, not merely a wording issue.
minor comments (6)
  1. [§4.2] The text refers to 'Table X' instead of 'Table 5' for the recoverability results. Please fix the cross-reference.
  2. [§3.1] There is a typo: 'Lets us mention' should be 'Let us mention.'
  3. [Table 1 and §3.3] The text mentions ALT in the discussion of liver enzymes, but Table 1 lists ALP, AST, and GGT without ALT. Also, the normal range for 'RNA Viral Load' is given as '50–Millions,' which is not a standard reference range; please clarify units and thresholds.
  4. [§3.3 and Table 5] The variable is called 'Continent of Birth' in the text and Table 2, but Table 5 abbreviates it as 'CoB' and labels it 'Country of Birth.' Please make the terminology consistent.
  5. [§3.4 and §3.5] Minor reproducibility issues: the MLP architecture and TabNet hyperparameters are not fully specified; no random seeds or data/code availability statement are given. I recommend adding these details in a reproducibility appendix.
  6. [References] References [18] and [19] appear to be the same paper (Shickel et al., 'Deep EHR: A Survey...') with overlapping details. Please merge or remove the duplicate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark; recoverability and aware/unaware comparisons are direct measurements, not reductions to inputs.

full rationale

The paper does not derive any target result from its own assumptions by construction. The central claims are empirical: Table 3 reports measured macro-F1 for demographic-aware vs demographic-unaware models across six algorithms, and Table 5 reports measured XGBoost recoverability of gender, age, race, and continent of birth from non-demographic inputs. These are trained-classifier evaluations under stratified k-fold cross-validation; no parameter is fitted to make a particular outcome true, and no equation equates a prediction with an input. The recoverability experiment's 'Lab Tests Only' configuration is defined as 'using all available laboratory and social measurements while excluding any demographic variables' (Section 4.2), so social features such as sexual orientation, commercial sex worker status, and route of infection are included; this weakens the paper's interpretation that gender/age are recoverable from laboratory data alone, but that is an experimental-design and framing flaw, not circular reasoning. There are also no load-bearing self-citations: all cited prior works are external references, and no uniqueness theorem or ansatz is imported from the authors' own prior work to force a choice. The aware/unaware gap is presented as an observed result with a post-hoc explanatory narrative, not as a quantity defined to equal the recoverability measurement. Therefore the paper's derivation chain is self-contained and no circular step is present.

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

The central feature-ablation result depends on the assumption that 'unaware' inputs contain no demographic proxies, which is doubtful given the social features. It also depends on the cross-sectional pairing of labs and ICD labels being treated as a prediction task, and on the cohort of 'complete records' being representative. No new entities are introduced.

free parameters (6)
  • Number of folds in stratified k-fold cross-validation = not reported
    k is a user choice; without it, no variance estimate is possible.
  • MLP architecture and training hyperparameters = not reported
    Layers, dropout, and learning rate are selected by the authors and can materially change F1.
  • TabNet hyperparameters = not reported
    Number of decision steps, attention size, and batch size are chosen by hand.
  • Random Forest tree count = 10 trees
    Small ensemble chosen without tuning; affects variance and central comparison.
  • LR inverse regularization strength C = 1.0
    Chosen without tuning.
  • Decision threshold for multi-label outputs = not reported (default 0.5 implied)
    Macro F1 depends on threshold; no threshold optimization or reporting is given.
assumptions (6)
  • domain assumption ICD-10 coded diagnoses in the hospital EHR are accurate ground truth for comorbidity labels.
    Used in Section 3.3 to define y_i; coding errors or incomplete specialty records would bias F1.
  • domain assumption The cohort of about 2,200 patients with complete records is representative of the HIV outpatient population.
    Section 3.3; no analysis of patients excluded for incomplete records, so selection bias cannot be assessed.
  • domain assumption Laboratory measurements and comorbidity diagnoses are temporally compatible for a prediction task.
    Paper frames 'proactive' prediction but Sections 3.1 and 3.3 specify no temporal split; if labs postdate diagnoses, the setup leaks outcome information.
  • domain assumption Social features used in the demographic-unaware arm carry no demographic or proxy information.
    Section 3.1: unaware models use laboratory and social features; sexual orientation, route of infection, and commercial sex worker status correlate with gender and age, so the ablation is not clean.
  • domain assumption Macro-F1 equally weighting 12 ICD categories is an appropriate performance measure.
    Section 3.4; with Z00-Z99 at 87%, equal weighting may overstate or understate clinical utility.
  • standard math Standard implementations of the cited ML algorithms (scikit-learn, PyTorch TabNet) are correct.
    Section 3.5; no custom derivation is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Proactive HIV Care: AI-Based Comorbidity Prediction from Routine EHR Data." pith.science (2026). https://pith.science/paper/SGJSRO5E

@misc{pith2026250820133,
  author       = {Pith},
  title        = {Pith review of: Proactive HIV Care: AI-Based Comorbidity Prediction from Routine EHR Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SGJSRO5E}},
  note         = {Machine review of arXiv:2508.20133}
}
read the original abstract

People living with HIV face a high burden of comorbidities, yet early detection is often limited by symptom-driven screening. We evaluate the potential of AI to predict multiple comorbidities from routinely collected Electronic Health Records. Using data from 2,200 HIV-positive patients in South East London, comprising 30 laboratory markers and 7 demographic/social attributes, we compare demographic-aware models (which use both laboratory/social variables and demographic information as input) against demographic-unaware models (which exclude all demographic information). Across all methods, demographic-aware models consistently outperformed unaware counterparts. Demographic recoverability experiments revealed that gender and age can be accurately inferred from laboratory data, underscoring both the predictive value and fairness considerations of demographic features. These findings show that combining demographic and laboratory data can improve automated, multi-label comorbidity prediction in HIV care, while raising important questions about bias and interpretability in clinical AI.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages

  1. [1]

    Arik and Tomas Pfister

    Sercan O. Arik and Tomas Pfister. Tabnet: Attentive in- terpretable tabular learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(8):6679–6687,

    work page

  2. [2]

    Bates, Suchi Saria, Lucila Ohno-Machado, Nigam Shah, and Gabriel Escobar

    David W. Bates, Suchi Saria, Lucila Ohno-Machado, Nigam Shah, and Gabriel Escobar. Big data in health care: using analytics to identify and manage high-risk and high-cost pa- tients. Health Affairs, 33(7):1123–1131, 2018. 2

    work page 2018

  3. [3]

    Random forests

    Leo Breiman. Random forests. Machine learning, 45:5–32,

    work page

  4. [4]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794. ACM, 2016. 3

    work page 2016

  5. [5]

    Recent advances and clin- ical applications of deep learning in medical image analysis

    Xuxin Chen, Ximin Wang, Ke Zhang, Kar-Ming Fung, Theresa C Thai, Kathleen Moore, Robert S Mannel, Hong Liu, Bin Zheng, and Yuchen Qiu. Recent advances and clin- ical applications of deep learning in medical image analysis. Medical image analysis, 79:102444, 2022. 2

    work page 2022

  6. [6]

    Chronic kidney disease stage identification in hiv infected patients using machine learning

    Asfakahemad Y Darveshwala, DheerajKumar Singh, and Yassir Farooqui. Chronic kidney disease stage identification in hiv infected patients using machine learning. In 2021 5th International Conference on Computing Methodologies and Communication (ICCMC), pages 1509–1514. IEEE, 2021. 2

    work page 2021

  7. [7]

    Metrics for multi-class classification: an overview

    Margherita Grandini, Enrico Bagli, and Giorgio Visani. Metrics for multi-class classification: an overview. arXiv preprint arXiv:2008.05756, 2020. 5

    arXiv 2008

  8. [8]

    Premature age-related comorbidi- ties among hiv-infected persons compared with the gen- eral population

    Giovanni Guaraldi et al. Premature age-related comorbidi- ties among hiv-infected persons compared with the gen- eral population. Clinical Infectious Diseases, 53(11):1120– 1126, 2011. HIV-infected patients show higher prevalence of non-infectious comorbidities (NICMs) compared to matched controls :contentReference[oaicite:2]index=2. 1

    work page 2011

Show all 23 references
  1. [9]

    Guaraldi et al

    G. Guaraldi et al. Cost of noninfectious comorbidities in patients with hiv. ClinicoEconomics and Outcomes Research, 5:481–488, 2013. Examines direct costs associ- ated with NICM care in HIV-infected patients :contentRef- erence[oaicite:3]index=3. 1

    2013

  2. [10]

    Global aids update 2023, 2023

    Joint United Nations Programme on HIV/AIDS (UN- AIDS). Global aids update 2023, 2023. Available on- line via UNAIDS global report summary :contentRefer- ence[oaicite:1]index=1. 1

    2023

  3. [11]

    Light- gbm: A highly efficient gradient boosting decision tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Light- gbm: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems (NeurIPS), pages 3146–3154, 2017. 3

    2017

  4. [12]

    Gordon, and Aldo A

    Matthieu Komorowski, Leo Anthony Celi, Omar Badawi, Anthony C. Gordon, and Aldo A. Faisal. The artificial intel- ligence clinician learns optimal treatment strategies for sep- sis in intensive care. Nature Medicine, 24(11):1716–1720,

  5. [13]

    An efficient approach to estimate the risk of coronary artery disease for people living with hiv us- ing machine-learning-based retinal image analysis

    Grace Lui, Ho Sang Leung, Jack Lee, Chun Kwok Wong, Xinxin Li, Mary Ho, Vivian Wong, Timothy Li, Tracy Ho, Yin Yan Chan, et al. An efficient approach to estimate the risk of coronary artery disease for people living with hiv us- ing machine-learning-based retinal image analysi...

    2023

  6. [14]

    Applied Logistic Regression Analysis

    Scott Menard. Applied Logistic Regression Analysis. SAGE Publications, 2002. 2

    2002

  7. [15]

    Kidd, and Joel T

    Riccardo Miotto, Fei Li, Brian A. Kidd, and Joel T. Dud- ley. Deep patient: An unsupervised representation to pre- dict the future of patients from the electronic health records. Scientific Reports, 6:26094, 2016. 2

    2016

  8. [16]

    Dai, et al

    Alvin Rajkomar, Eyal Oren, Kai Chen, Andrew M. Dai, et al. Scalable and accurate deep learning with electronic health records. npj Digital Medicine, 1(1):18, 2018. 2

    2018

  9. [17]

    Stratification for multi-label data

    Konstantinos Sechidis, Grigorios Tsoumakas, and Ioan- nis Vlahavas. Stratification for multi-label data. In Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD) Workshop on Knowledge Discovery and Data Mining, pages 145–15...

    2011

  10. [18]

    Deep ehr: a survey of recent advances in deep learning techniques for electronic health record (ehr) analysis

    Benjamin Shickel, Patrick James Tighe, Azra Bihorac, and Parisa Rashidi. Deep ehr: a survey of recent advances in deep learning techniques for electronic health record (ehr) analysis. IEEE journal of biomedical and health informatics, 22(5):1589–1604, 2017. 2

    2017

  11. [19]

    Deep ehr: A survey of recent advances in deep learning techniques for electronic health record (ehr) analy- sis

    Benjamin Shickel, Patrick J Tighe, Azra Bihorac, and Parisa Rashidi. Deep ehr: A survey of recent advances in deep learning techniques for electronic health record (ehr) analy- sis. IEEE Journal of Biomedical and Health Informatics, 22 (5):1589–1604, 2018. 2

    2018

  12. [20]

    Investigating the impact of data normalization on classification performance

    Dalwinder Singh and Birmohan Singh. Investigating the impact of data normalization on classification performance. Applied Soft Computing, 97:105524, 2020. 5

    2020

  13. [21]

    Eric J. Topol. High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine, 25(1): 44–56, 2019. 2

    2019

  14. [22]

    An interpretable mortality prediction model for covid-19 patients

    Liang Yan, Haibo Zhang, Jorge Goncalves, Yuan-Ting Xiao, Meng Wang, Yali Guo, Chuan Sun, Xuan Tang, Lin Jin, and Ming Zhang. An interpretable mortality prediction model for covid-19 patients. Nature Machine Intelligence, 2(5):283– 288, 2020. 2

    2020

  15. [23]

    Machine learning-based prognostic prediction for hospitalized hiv/aids patients with cryptococcus infection in guangxi, china

    Baili Zhan, Wudi Wei, Zhiman Xie, Sirun Meng, Xiuli Bao, Xiaotao He, Xiaoting Xie, Meng Zhang, Li Ye, Junjun Jiang, et al. Machine learning-based prognostic prediction for hospitalized hiv/aids patients with cryptococcus infection in guangxi, china. BMC Infectious Diseases, 24...

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.