Multi-Source COVID-19 Detection via Variance Risk Extrapolation

Hao Chen; Jilan Xu; Junlin Hou; Qingqiu Li; Rui Feng; Runtian Yuan; Yuejie Zhang

REVIEW 4 major objections 5 minor 18 references

Multi-Source COVID-19 Detection via Variance Risk Extrapolation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that VREx domain-aligned pretraining followed by Mixup fine-tuning reaches a 0.96 average macro F1 for COVID-19 detection on chest CT scans from four hospitals.

desk verdict Thin challenge report: the 0.96 F1 is not evidence for VREx/Mixup without an ERM baseline or ablations. read the letter →

arxiv 2506.23208 v1 pith:DSNMFRWH submitted 2025-06-29 eess.IV cs.CV

Runtian Yuan , Qingqiu Li , Junlin Hou , Jilan Xu , Yuejie Zhang , Rui Feng , Hao Chen This is my paper

classification eess.IVcs.CV

keywords COVID-19detectionchestCTdomaingeneralizationvarianceriskextrapolationMixupaugmentationmulti-sourcemedicalimagingmacroF1

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 tackles a practical obstacle in automated COVID-19 triage: chest CT scans from different hospitals look different because scanners, imaging protocols, and patient populations vary, and a model trained on one center can fail on another. Its proposed recipe trains in two stages: first pretrain with Variance Risk Extrapolation (VREx), an objective that minimizes both the average classification loss and the variance of that loss across the four known source domains, pushing the encoder toward features that behave consistently across hospitals; then fine-tune with Mixup, which creates new training examples by interpolating pairs of images and their labels, including pairs from different domains, to flatten decision boundaries. On the challenge validation set, the paper reports an average macro F1 (the mean of the per-class F1 scores) of 0.96 across the four sources. The significance, if the result holds, is that a labeled multi-center dataset plus a variance penalty and input mixing can yield a detector that does not lean on center-specific artifacts.

What carries the argument

The load-bearing object is the VREx training objective, written by the paper as $\min_\theta \left( \frac{1}{n}\sum_{i=1}^n \mathcal{L}_i(\theta) + \lambda \operatorname{Var}[\mathcal{L}_i(\theta)] \right)$, where $\mathcal{L}_i$ is the classification loss on domain $i$ and $\lambda$ balances average risk against its variance; this term is what enforces cross-domain consistency during pretraining. The second half of the mechanism is Mixup, which generates virtual examples $\tilde{x} = \lambda x_i + (1-\lambda)x_j$ and $\tilde{y} = \lambda y_i + (1-\lambda)y_j$ from pairs of training samples, potentially across domains, to regularize fine-tuning. The encoder on which both stages run is not specified, so the contribution is carried by the two training losses rather than by a new architecture.

What would settle it

Train the same encoder on the same data with ordinary loss minimization in place of the VREx penalty and with Mixup disabled; if the average macro F1 stays at 0.96, or if none of the four per-domain scores changes, the proposed stages are not responsible for the reported generalization.

Watch

Extended reading notes

Core claim

The central claim is that the paper's two-stage framework—VREx domain-aligned pretraining followed by Mixup domain-mixed fine-tuning—classifies COVID-19 versus non-COVID chest CT scans with an average macro F1 of 0.96 when evaluated across four distinct hospital domains. In Stage 1, the model minimizes a weighted combination of the mean per-domain empirical risk and the variance of those risks, so no single institution's loss pattern is allowed to dominate. In Stage 2, Mixup interpolates inputs and labels, and the paper argues this smooths decision boundaries and reduces overfitting to domain-specific features. The 0.96 macro F1 on the held-out validation set is presented as evidence that the model has captured domain-invariant representations rather than memorizing center-specific cues.

Load-bearing premise

The load-bearing assumption is that the reported 0.96 macro F1 comes from the VREx and Mixup stages rather than from the unspecified pretrained encoder or the particular validation set, and the paper gives no baseline or ablation that would separate those causes.

Editorial extensions

If this is right

A hospital with labeled local data can run the same two-stage recipe and expect a detector that carries over to other centers without retraining on their labels, since neither stage needs target-domain data.
The variance penalty gives a concrete way to stop models latching onto scanner- and protocol-specific cues whenever source domains are known during training.
Cross-domain Mixup during fine-tuning provides a route to smoother decision boundaries on small per-center datasets, as the 1,124 training scans here require.
The 0.96 aggregate macro F1 becomes a benchmark number that future multi-source COVID-19 detectors can be compared against on the same validation set.

Reading between the lines

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

Per-domain F1 scores would show whether the 0.96 average is uniform across the four hospitals or carried by one; the paper reports only the aggregate, but macro averaging makes this check straightforward.
Applying the VREx variance penalty during fine-tuning as well as pretraining, or weighting the penalty per domain, are natural variants the paper leaves untested.
Since the method needs only known source-domain labels, the same recipe could transfer to other multi-source medical imaging tasks, such as pneumonia or lung-nodule classification, whenever hospital of origin is recorded.

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.

Desk Editor's Note

read the letter

Short version: this is a cleanly written two-page challenge write-up, but the central claim is unsupported. The authors say VREx pretraining plus Mixup fine-tuning gives 0.96 macro F1 across four hospitals, and they attribute that to the method. There is no ERM baseline, no ablation, no per-domain numbers, no seed variance, and no architecture or hyperparameter details. So on the evidence presented, the number could be mostly the pretrained CMC encoder (which is only cited, not specified) or luck on a 308-scan validation set.

What's actually decent: the paper describes the two-stage pipeline clearly, states the VREx loss (Eq 1) and Mixup (Eq 2) correctly, and doesn't claim a new algorithm. It's a straightforward combination of known techniques, which is fine for a challenge report. The writing is honest about being a challenge solution. That's about all.

The soft spots are the load-bearing ones. The stress-test note is right: the encoder is 'the CMC network' from Refs [4-7], which are COVID-19 CT methods, but the model is never named, its pretraining data is not described, and it is not released. If that encoder was already trained on COVID CT, the marginal gain from VREx and Mixup could be small or negative. Without a plain ERM run, you can't tell. Also, validation has only 308 scans with 128 positives across four domains, so one small domain can move the average; an aggregate 0.96 is under-specified. The paper also gives no lambda values, Mixup alpha, optimizer, or epoch counts, so the experiment is not reproducible.

The conclusion 'demonstrating strong generalization' is not supported. The result might be true, but the evidence in the paper is a single scalar. I'd want the baseline, ablations, per-domain table, and code before believing the attribution.

Who this is for: someone tracking the PHAROS-AFE-AIMI challenge or looking for a compact example of VREx in medical imaging. It's not a research paper that advances domain generalization.

Recommendation: I'd desk-reject this as a regular paper. If the venue has a workshop/challenge track, it could be accepted as a system report, but it should at least add an ERM comparison. Right now it doesn't deserve serious referee time.

Referee Report

4 major / 5 minor

Summary. This manuscript reports a challenge solution for the Multi-Source COVID-19 Detection Challenge, which requires binary classification of chest CT scans into COVID and Non-COVID categories across four hospital sources. The proposed method is a two-stage framework: Stage 1 uses Variance Risk Extrapolation (VREx) as a domain-aligned pretraining objective that minimizes the variance of empirical risks across source domains, and Stage 2 fine-tunes the model with Mixup data augmentation. The paper reports an average macro F1 score of 0.96 on the validation set and claims that this demonstrates strong generalization. The manuscript is very short, contains no comparison to empirical risk minimization (ERM), no ablations, no per-domain results, no repeated-run statistics, and no hyperparameter details.

Significance. If the reported result is reproducible and the attribution to VREx and Mixup is correct, the manuscript would demonstrate a practically useful recipe for cross-institutional COVID-19 CT detection, which is a relevant clinical application. However, the paper currently does not establish that the two proposed components cause the improvement over the unspecified encoder or over standard training, and both VREx and Mixup are well-known techniques. The main value of the paper is therefore as a challenge-solution report rather than as a methodological advance. The manuscript is concise and the problem is clearly stated, but the lack of rigorous evaluation limits its significance.

major comments (4)

[Section 4] The central claim that VREx pretraining and Mixup fine-tuning yield an average macro F1 of 0.96 is unsupported because Section 4 reports only a single scalar with no ERM baseline, no lambda=0 control, no VREx-only or Mixup-only ablation, no per-domain macro F1, and no repeated-run statistics. Add these experiments, including seed variance and a per-domain breakdown, or the attribution of the score to the proposed components does not follow.
[Section 2, Stage 1] The encoder is never specified: the text says the framework builds on the CMC network [4-7], but those references are COVID-19 CT classification methods, and the architecture, pretraining data, and initialization are omitted. Without an ERM baseline using the same encoder, or an ablation that removes the VREx term from Eq. (1), a reader cannot determine whether the 0.96 F1 comes from VREx and Mixup or from a pretrained representation that already separates COVID from Non-COVID.
[Section 3] The validation set composition in Table 1 is given only in aggregate: 180 Non-COVID vs 128 COVID, with no per-domain counts. Because the reported macro F1 averages four sources and some sources may have very few positive cases, the aggregate number is unstable and under-specified; report the per-domain distribution and interval estimates such as bootstrap confidence intervals or per-seed results.
[Section 2] The experimental configuration is not reproducible: the value of lambda in Eq. (1), the Mixup parameter in Eq. (2), optimizer, learning rate, batch size, number of epochs, image preprocessing, and evaluation protocol are all missing. Provide a complete hyperparameter table, or state that the challenge protocol fixes these values with a citation.

minor comments (5)

[Eqs. (1) and (2)] The symbol lambda is used both for the VREx penalty coefficient in Eq. (1) and for the Mixup interpolation weight in Eq. (2); using different symbols (e.g., beta for Mixup) would avoid ambiguity.
[Figure 1] Figure 1 appears as a low-resolution diagram with garbled or unreadable labels in the provided manuscript; redraw it with clear text for the loss terms and stage division.
[Section 2] The phrase 'PHAROS-AFE-AIMI' is not defined; state the full name of the challenge and provide a reference or URL.
[Abstract] The abstract says the result 'demonstrating strong generalization,' but only a validation set is used; no test set is discussed, so generalization beyond the validation split is not actually demonstrated.
[References] Reference [18] is cited as an arXiv preprint; the published ICLR 2018 version of the Mixup paper should be cited if available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports a directly measured validation F1; neither VREx/Mixup attribution nor the CMC self-citation forces the result by construction.

full rationale

The paper contains no derivation chain whose output is equivalent to its input. Section 4 reports an average macro F1 of 0.96 as a direct measurement on the validation set, not as a quantity fitted from, or defined in terms of, the training objective. Equation (1) is the standard VREx objective and Eq. (2) is standard Mixup; neither equation implicitly contains the reported F1, and no parameter is tuned so that the F1 follows by construction. The reference to the authors' earlier CMC network [4-7] is used as an architectural foundation, but the paper's central claim is an empirical score, not a theorem derived from that citation. The absence of an ERM baseline, ablations, per-domain statistics, or repeated-run variance weakens the causal attribution of the score to VREx and Mixup, but that is a lack of experimental support rather than circularity. Under the hard rules, concerns about attribution without controlled comparison do not constitute a circular step because no specific equation or fitted quantity reduces to its own input. The honest finding is therefore no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new theory or entities, but its empirical claims depend on several assumptions about VREx and Mixup behavior that are not tested here. All loss coefficients and training hyperparameters are undisclosed, and the encoder is not specified.

free parameters (3)

VREx variance penalty coefficient lambda
Introduced in Eq. (1) as the regularization weight balancing empirical risk and its variance; its value is not reported.
Mixup interpolation weight or distribution parameter
Appears in Eq. (2) as the interpolation coefficient lambda (or the parameter of the underlying Beta distribution); the paper does not report the value or distribution used.
Training hyperparameters (learning rate, batch size, epochs, optimizer)
Central to any deep learning result, yet none are disclosed in the paper.

assumptions (5)

domain assumption The VREx objective in Eq. (1) encourages domain-invariant representations that generalize to unseen domains.
Adopted from Krueger et al. [17] without testing on this dataset; the paper relies on this to interpret the F1 score as evidence of domain alignment.
domain assumption Mixup in Eq. (2) improves robustness and generalization.
Adopted from Zhang et al. [18]; no ablation supports this assumption in the reported experiment.
domain assumption The four training domains adequately represent the relevant sources of domain shift.
The challenge defines four hospitals; the paper assumes that minimizing risk variance across them yields a generalizable model, without reporting per-domain prevalence or characteristics.
domain assumption The validation set is an unbiased sample from the target distribution and the labels are correct.
Section 3 provides only aggregate counts; there is no evidence of the per-domain validation composition or label quality.
domain assumption The unspecified encoder has sufficient capacity and initialization to support the classification task.
Figure 1 shows a generic encoder and classifier, but the backbone and pretraining details are not given, so the contribution of the architecture to the result is unknown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Source COVID-19 Detection via Variance Risk Extrapolation." pith.science (2026). https://pith.science/paper/DSNMFRWH

@misc{pith2026250623208,
  author       = {Pith},
  title        = {Pith review of: Multi-Source COVID-19 Detection via Variance Risk Extrapolation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DSNMFRWH}},
  note         = {Machine review of arXiv:2506.23208}
}

read the original abstract

We present our solution for the Multi-Source COVID-19 Detection Challenge, which aims to classify chest CT scans into COVID and Non-COVID categories across data collected from four distinct hospitals and medical centers. A major challenge in this task lies in the domain shift caused by variations in imaging protocols, scanners, and patient populations across institutions. To enhance the cross-domain generalization of our model, we incorporate Variance Risk Extrapolation (VREx) into the training process. VREx encourages the model to maintain consistent performance across multiple source domains by explicitly minimizing the variance of empirical risks across environments. This regularization strategy reduces overfitting to center-specific features and promotes learning of domain-invariant representations. We further apply Mixup data augmentation to improve generalization and robustness. Mixup interpolates both the inputs and labels of randomly selected pairs of training samples, encouraging the model to behave linearly between examples and enhancing its resilience to noise and limited data. Our method achieves an average macro F1 score of 0.96 across the four sources on the validation set, demonstrating strong generalization.

Figures

Figures reproduced from arXiv: 2506.23208 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 4 canonical work pages

[1]

A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis

Anastasios Arsenos, Dimitrios Kollias, and Stefanos Kol- lias. A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis. In 2022 IEEE 14th Im- age, Video, and Multidimensional Signal Processing Work- shop (IVMSP), page 1–5. IEEE, 2022. 1

2022
[2]

Data-driven covid-19 detection through medical imaging

Anastasios Arsenos, Andjoli Davidhi, Dimitrios Kollias, Panos Prassopoulos, and Stefanos Kollias. Data-driven covid-19 detection through medical imaging. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), page 1–5. IEEE, 2023

2023
[3]

Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system

Demetris Gerogiannis, Anastasios Arsenos, Dimitrios Kol- lias, Dimitris Nikitopoulos, and Stefanos Kollias. Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system. In 2024 IEEE In- ternational Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2024. 1

2024
[4]

Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis

Junlin Hou, Jilan Xu, Rui Feng, Yuejie Zhang, Fei Shan, and Weiya Shi. Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 454–461,

work page
[5]

Periphery-aware covid-19 diagnosis with contrastive repre- sentation enhancement

Junlin Hou, Jilan Xu, Longquan Jiang, Shanshan Du, Rui Feng, Yuejie Zhang, Fei Shan, and Xiangyang Xue. Periphery-aware covid-19 diagnosis with contrastive repre- sentation enhancement. Pattern Recognition, 118:108005, 2021

2021
[6]

Cmc v2: Towards more accu- rate covid-19 detection with discriminative video priors

Junlin Hou, Jilan Xu, Nan Zhang, Yi Wang, Yuejie Zhang, Xiaobo Zhang, and Rui Feng. Cmc v2: Towards more accu- rate covid-19 detection with discriminative video priors. In European Conference on Computer Vision, pages 485–499. Springer, 2022

work page 2022
[7]

Boosting covid-19 severity detec- tion with infection-aware contrastive mixup classification

Junlin Hou, Jilan Xu, Nan Zhang, Yuejie Zhang, Xiaobo Zhang, and Rui Feng. Boosting covid-19 severity detec- tion with infection-aware contrastive mixup classification. In European Conference on Computer Vision, pages 537–551. Springer, 2022. 1

work page 2022
[8]

Deep neural archi- tectures for prediction in healthcare

Dimitrios Kollias, Athanasios Tagaris, Andreas Stafylopatis, Stefanos Kollias, and Georgios Tagaris. Deep neural archi- tectures for prediction in healthcare. Complex & Intelligent Systems, 4(2):119–131, 2018. 1

2018

Show all 18 references

[9]

Deep transparent prediction through latent represen- tation analysis

Dimitrios Kollias, N Bouas, Y Vlaxos, V Brillakis, M Se- feris, Ilianna Kollia, Levon Sukissian, James Wingate, and S Kollias. Deep transparent prediction through latent represen- tation analysis. arXiv preprint arXiv:2009.07044, 2020

2009 arXiv
[10]

Transpar- ent adaptation in deep medical image diagnosis

Dimitris Kollias, Y Vlaxos, M Seferis, Ilianna Kollia, Levon Sukissian, James Wingate, and Stefanos D Kollias. Transpar- ent adaptation in deep medical image diagnosis. In TAILOR, page 251–267, 2020

2020
[11]

Mia-cov19d: Covid-19 detection through 3-d chest ct image analysis

Dimitrios Kollias, Anastasios Arsenos, Levon Soukissian, and Stefanos Kollias. Mia-cov19d: Covid-19 detection through 3-d chest ct image analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision , page 537–544, 2021

2021
[12]

Ai-mia: Covid-19 detection and severity analysis through medical imaging

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Ai-mia: Covid-19 detection and severity analysis through medical imaging. In European Conference on Computer Vi- sion, page 677–690. Springer, 2022

2022
[13]

Ai-enabled analysis of 3-d ct scans for diagnosis of covid-19 & its severity

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Ai-enabled analysis of 3-d ct scans for diagnosis of covid-19 & its severity. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICAS- SPW), page 1–5. IEEE, 2023

2023
[14]

A deep neural architecture for harmonizing 3-d input data analysis and decision making in medical imaging

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. A deep neural architecture for harmonizing 3-d input data analysis and decision making in medical imaging. Neuro- computing, 542:126244, 2023

2023
[15]

Domain adaptation explainability & fairness in ai for medical image analysis: Diagnosis of covid-19 based on 3-d chest ct-scans

Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kol- lias. Domain adaptation explainability & fairness in ai for medical image analysis: Diagnosis of covid-19 based on 3-d chest ct-scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

2024
[16]

Sam2clip2sam: Vision language model for segmentation of 3d ct scans for covid-19 detection

Dimitrios Kollias, Anastasios Arsenos, James Wingate, and Stefanos Kollias. Sam2clip2sam: Vision language model for segmentation of 3d ct scans for covid-19 detection. arXiv preprint arXiv:2407.15728, 2024. 1

2024 arXiv
[17]

Out-of-distribution general- ization via risk extrapolation (rex)

David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution general- ization via risk extrapolation (rex). In International confer- ence on machine learning, pages 5815–5826. PMLR, 2021. 1

2021
[18]

mixup: Beyond empirical risk minimiza- tion

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion. arXiv preprint arXiv:1710.09412, 2017. 2 3

2017 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis

Anastasios Arsenos, Dimitrios Kollias, and Stefanos Kol- lias. A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis. In 2022 IEEE 14th Im- age, Video, and Multidimensional Signal Processing Work- shop (IVMSP), page 1–5. IEEE, 2022. 1

2022

[2] [2]

Data-driven covid-19 detection through medical imaging

Anastasios Arsenos, Andjoli Davidhi, Dimitrios Kollias, Panos Prassopoulos, and Stefanos Kollias. Data-driven covid-19 detection through medical imaging. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), page 1–5. IEEE, 2023

2023

[3] [3]

Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system

Demetris Gerogiannis, Anastasios Arsenos, Dimitrios Kol- lias, Dimitris Nikitopoulos, and Stefanos Kollias. Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system. In 2024 IEEE In- ternational Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2024. 1

2024

[4] [4]

Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis

Junlin Hou, Jilan Xu, Rui Feng, Yuejie Zhang, Fei Shan, and Weiya Shi. Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 454–461,

work page

[5] [5]

Periphery-aware covid-19 diagnosis with contrastive repre- sentation enhancement

Junlin Hou, Jilan Xu, Longquan Jiang, Shanshan Du, Rui Feng, Yuejie Zhang, Fei Shan, and Xiangyang Xue. Periphery-aware covid-19 diagnosis with contrastive repre- sentation enhancement. Pattern Recognition, 118:108005, 2021

2021

[6] [6]

Cmc v2: Towards more accu- rate covid-19 detection with discriminative video priors

Junlin Hou, Jilan Xu, Nan Zhang, Yi Wang, Yuejie Zhang, Xiaobo Zhang, and Rui Feng. Cmc v2: Towards more accu- rate covid-19 detection with discriminative video priors. In European Conference on Computer Vision, pages 485–499. Springer, 2022

work page 2022

[7] [7]

Boosting covid-19 severity detec- tion with infection-aware contrastive mixup classification

Junlin Hou, Jilan Xu, Nan Zhang, Yuejie Zhang, Xiaobo Zhang, and Rui Feng. Boosting covid-19 severity detec- tion with infection-aware contrastive mixup classification. In European Conference on Computer Vision, pages 537–551. Springer, 2022. 1

work page 2022

[8] [8]

Deep neural archi- tectures for prediction in healthcare

Dimitrios Kollias, Athanasios Tagaris, Andreas Stafylopatis, Stefanos Kollias, and Georgios Tagaris. Deep neural archi- tectures for prediction in healthcare. Complex & Intelligent Systems, 4(2):119–131, 2018. 1

2018