HumekaFL: Automated Detection of Neonatal Asphyxia Using Federated Learning

Assane Gueye; Bereket Retta; Blessed Guda; Carlee Joe-Wong; Gladys Inabeza; Pamely Zantou

REVIEW 3 major objections 6 minor 9 references

HumekaFL: Automated Detection of Neonatal Asphyxia Using Federated Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A federated SVM trained on infant cries detects birth asphyxia on-device with sensitivity 0.91 and specificity 0.96, beating the centralized baseline in the paper it compares against.

desk verdict Useful FL prototype for cry-based asphyxia detection, but the headline claim that FedSVM beats centralized models is not supported by the experiments. read the letter →

arxiv 2412.01167 v1 pith:KB5VAAHT submitted 2024-12-02 cs.LG eess.AS

Pamely Zantou , Blessed Guda , Bereket Retta , Gladys Inabeza , Carlee Joe-Wong , Assane Gueye This is my paper · ORCID

classification cs.LGeess.AS

keywords federatedlearningbirthasphyxiainfantcryanalysissupportvectormachineMFCCFedAvgmobilehealthapplicationneonatalmortality

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 argues that the privacy barrier to machine-learning diagnosis of birth asphyxia can be removed without sacrificing accuracy. It proposes HumekaFL, a federated-learning mobile pipeline in which ten hospital silos train a shared linear-kernel SVM on MFCC features extracted from one-second cry recordings, with only model weights shared through FedAvg. On the Baby Chillanto dataset, the resulting FedSVM model reports sensitivity 0.91, specificity 0.96, and unweighted average recall 0.93, and the federated run reaches 95.88% average accuracy over 50 rounds. If the comparison holds, hospitals can adopt early asphyxia screening while keeping newborn cry data on premises. The paper's own comparison, however, bundles federation together with three other changes, so the cause of the improvement needs disentangling.

What carries the argument

The load-bearing object is FedSVM: a linear-kernel support vector machine trained locally with Adam on each of ten clients and aggregated with Federated Averaging. Each one-second cry clip is converted into 40 Mel-Frequency Cepstral Coefficients, augmented with tanh distortion and room reverberation, and pruned with a Random Forest feature selector. The mechanism links four parts: local regularized hinge-loss optimization on each silo; FedAvg weight averaging proportional to local data volume, allowing training without raw data leaving a hospital; the small SVM parameter footprint, which keeps the model deployable on commodity mobile hardware; and the inference chain of voice activity detection, Butterworth band-pass filtering, MFCC extraction, and on-device classification.

What would settle it

Retrain the same pipeline with augmentation, hyperparameters, and MFCC selection held fixed, but replace federated averaging with one centralized training run on pooled data; if the centralized run matches or exceeds 0.93 UAR and 95.88% accuracy, the claim that federated training outperforms centralized training is falsified.

Watch

Extended reading notes

Core claim

HumekaFL's central discovery claim is that a federated SVM can detect birth asphyxia from cry sounds with performance above published centralized approaches that use the same Baby Chillanto dataset. The authors train a linear-kernel SVM that minimizes the regularized hinge-loss objective $F(w) = \frac{\lambda}{2}\|w\|_2^2 + \frac{1}{n}\sum_{i=1}^n \max(0, 1 - w^T x_i y_i)$, optimized locally with Adam on ten simulated hospital silos and aggregated by FedAvg as $w^{t+1}_{glob} = \sum_{k \in S_t} \frac{n_k}{n} w^{t+1}_k$. They report FedSVM sensitivity 0.91, specificity 0.96, and UAR 0.93 against the centralized SVM baseline's sensitivity 0.85, specificity 0.89, and UAR 0.865, plus 95.88% average accuracy over 50 federated rounds, leading them to conclude that the federated pipeline outperforms centralized SVM and NN-based models on the Baby Chillanto dataset.

Load-bearing premise

The load-bearing premise is that FedSVM's better numbers than the published centralized SVM come from the federated training setup, even though the paper's own comparison also changed data augmentation, hyperparameter tuning, and MFCC feature selection.

Editorial extensions

If this is right

Hospitals can participate in collaborative training of an asphyxia detector without exporting newborn cry recordings, addressing the privacy concern that blocks centralized ML adoption in African healthcare settings.
A working 10-second on-device screening workflow (VAD, filtering, MFCC extraction, SVM inference) can be run by health workers or caregivers with no machine-learning training, on commodity hardware.
The reported 95.88% average federated accuracy after 50 rounds indicates convergence is achievable with ten silos and five local epochs, though the authors note results are preliminary until tested with physical healthcare clients.
The comparison against published SVM and neural-network methods supports the claim that small, resource-lean models can compete with deep networks on this small cry dataset.

Reading between the lines

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

Because the comparison bundles federation with data augmentation, hyperparameter tuning, and MFCC feature selection, the reported accuracy gain cannot be attributed to federated learning alone; a centralized run with the same three enhancements would isolate the federation contribution.
If federation genuinely helps on this small benchmark, a likely mechanism is implicit regularization from constrained local updates and averaging, not privacy alone; this could be tested by ablating the number of silos and local epochs.
The Baby Chillanto data is Mexican, so transferring FedSVM to sub-Saharan settings would require collecting African cries, since MFCC distributions and recording conditions may shift with language, environment, and clinical practice.
The mobile application's field behavior could be tested prospectively against clinical APGAR assessment in hospitals; no such clinical validation is reported.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: this is a prototype paper, not a results paper. The authors build a federated SVM pipeline for cry-based asphyxia detection and wrap it in a mobile app for low-resource settings. That's a real contribution of the systems kind. But the abstract and conclusion claim FedSVM outperformed centralized SVM and NN baselines, and the experiments don't support that.

What's actually new: applying FedAvg to a linear SVM with MFCC features on Baby Chillanto is not novel individually, but the end-to-end architecture—silos, aggregation, on-device inference—is a useful reference for anyone building FL health tools in African settings. The authors also list limitations themselves: results are preliminary, they need African data, and differential privacy is deferred. That honesty counts.

Where it falls down: the comparison to Ubenwa in Table 1 is confounded. The text says the main differences are data augmentation, hyperparameter tuning, MFCC feature selection, and the federated setup. Without ablating those, you can't attribute the gain to FL. Worse, Table 1's numbers are not clearly labeled as coming from the federated model. The one number explicitly tied to FedSVM is 95.88% average training accuracy, which is not a held-out metric. So the central claim is untested. There are also no error bars, no data split details, no code.

Who should read it: people working on deployment prototypes for cry analysis or FL in low-resource health systems might get architectural ideas. Anyone evaluating the FL-vs-centralized question should not cite this.

Recommendation: I would not send this to a serious journal or conference in its current form. It needs a matched centralized baseline, held-out evaluation of the actual federated model, and an ablation separating augmentation/tuning/feature selection from the learning setup. A workshop that accepts systems demos might take it with the claims toned down. For peer review, I'd desk reject or require major revision; the evidence isn't there yet.

Referee Report

3 major / 6 minor

Summary. The paper proposes HumekaFL, a federated learning mobile application for detecting neonatal asphyxia from newborn cries. Using the Baby Chillanto dataset, MFCC features, tanh/room-reverberation data augmentation, Random Forest feature selection, and FedAvg over ten virtual silos, the authors train a linear SVM. They report a HumekaFL SVM with sensitivity 0.91, specificity 0.96, and UAR 0.93, which they compare to Ubenwa's published SVM numbers, and they additionally report a FedSVM average training accuracy of 95.88% over fifty communication rounds. The abstract and conclusion assert that the federated SVM pipeline outperforms centralized SVM and neural-network-based methods.

Significance. If properly established, the central claim would be valuable: a privacy-preserving, deployable FL system for cry-based asphyxia detection on commodity hardware, with a public dataset and a mobile application, would be a useful contribution to neonatal health in low-resource settings. The paper's strengths are its clear motivation, concrete system architecture, and explicit use of FedAvg with a small SVM model. However, the current experimental evidence does not isolate the federated training setup from other pipeline changes, so the headline claim that federated learning outperforms centralized training is untested. The comparison against Ubenwa is an external benchmark, not a controlled experiment.

major comments (3)

[Section 3, Table 1, Abstract, Conclusion] The central claim that FedSVM outperforms centralized SVM and NN-based methods is not supported by the reported experiments. The text explicitly lists four simultaneous differences between the HumekaFL SVM and Ubenwa's SVM: data augmentation, hyperparameter tuning, feature selection of MFCCs, and the federated training setup. No ablation or centralized version of the same pipeline is trained, so any of these differences, or a combination, could explain the performance gap. Furthermore, the Table 1 metrics are introduced as 'HumekaFL SVM' before the federated experiment is described, and the only numeric result explicitly attributed to FedSVM is the 95.88% average training accuracy in the second set of experiments. No held-out sensitivity, specificity, or UAR is reported for the federated model. The authors' own caveat that 'we need to conduct more experiments with physical healthcare clients to validate these preliminary results' is in tension with the categorical conclusion that FedSVM outperformed existing centralized models.
[Section 2, Section 3] The evaluation protocol is underspecified, which makes the reported numbers impossible to interpret or compare. The paper does not state how the Baby Chillanto recordings were divided into training and test sets, whether the split was stratified, whether data augmentation was applied before or after the split, how UAR was computed, or whether metrics are averaged across clients or communication rounds. If augmentation is applied before partitioning, the same original cry may appear in both training and test sets, which would inflate the reported metrics. Because Ubenwa's published numbers may have been obtained under a different evaluation protocol, the comparison in Table 1 is not meaningful without this information.
[Section 2, Data Pre-processing] The binary classification setup is unclear. The dataset description includes 1,049 healthy recordings, 870 deaf-infant recordings, and 340 asphyxia recordings, plus five types of cry signals, but after augmentation the paper reports only 1,521 normal and 1,028 asphyxiated data points. The paper should state which recordings were kept, how deaf and pain cries were labeled or excluded, and whether Ubenwa used the same subset. If the class definitions or data subsets differ, the Table 1 comparison may be comparing different tasks rather than different learning paradigms.

minor comments (6)

[Abstract and title] The word 'Asphyxia' is misspelled as 'Apshyxia' in the abstract and in the section heading; this should be corrected throughout.
[Section 2, Eq. (2)] Equation (2) for tanh distortion is not a standard expression and leaves the variables a, b, and G undefined; please provide the actual augmentation formula and parameter values used in the experiments.
[Section 2, Federated SVM] The paper uses Adam to solve the hinge-loss SVM objective but does not specify the local update rule or how the resulting SVM weight vectors are aggregated under FedAvg; since FedAvg is normally described for SGD-based updates, this should be clarified.
[Section 3, Figure 4] Figure 4 reports average training accuracy over communication rounds; adding a validation or test accuracy curve, or explaining why training accuracy is the reported metric, would substantially strengthen the empirical section.
[Section 2, System Architecture] The phrase 'cross-silo and centralized federated learning architecture' is ambiguous; please define the terminology or use a standard term such as 'centralized federated learning' with cross-silo clients.
[Section 3 and Conclusion] The conclusion mentions outperforming NN-based methods, but no neural network baseline is trained in the paper; if the comparison is to literature values, the specific systems and datasets should be cited and described.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported numbers are empirical benchmark results, not derived quantities; the main weakness is an uncontrolled comparison, not a self-referential derivation chain.

full rationale

No circular derivation chain is present. The FedSVM system uses standard SVM optimization (regularized empirical risk with hinge loss, Eqs. 3-4) and FedAvg aggregation, both defined independently of the target performance claim, with external citations for the FL algorithms. The reported metrics in Table 1 (sensitivity 0.91, specificity 0.96, UAR 0.93) are empirical comparisons against Ubenwa's published SVM, and the federated experiment reports a 95.88% average training accuracy; no fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no equation reduces to its own input by construction. The paper's own text lists four simultaneous differences from Ubenwa (data augmentation, hyperparameter tuning, MFCC feature selection, and the federated setup) and does not ablate the federated component, so the headline claim that federated learning outperforms centralized training is not experimentally established. That is a validity and attribution concern, not circularity, and the paper does include a preliminary-results caveat. Consequently, the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper's central claim rests on domain assumptions about label validity, feature sufficiency, augmentation preserving labels, FL convergence, and recording, along with several unreported hyperparameters. No code or full parameter set is provided.

free parameters (6)

SVM regularization coefficient lambda
Introduced in Eq. 3; value not reported; tuned by hand or search.
Number of MFCCs = 40
Chosen as feature representation; hand-selected without reported justification.
Tanh distortion gain G
Data augmentation parameter in Eq. 2; value not specified; affects all augmented samples.
Room reverberation filter settings
RIR convolution parameters not reported; affects augmented data.
Number of selected features after Random Forest
Feature selection step described but no count or list given.
Adam optimizer hyperparameters
Learning rate and other optimizer settings not reported for local SVM training.

assumptions (5)

domain assumption Baby Chillanto labels (asphyxia vs normal) are correct and clinically valid
The model is trained and evaluated on these labels; no independent verification is provided.
domain assumption MFCC features contain sufficient discriminative information for asphyxia detection
Inherited from prior cry-analysis literature [3,8]; assumed without validation on this dataset.
ad hoc to paper Data augmentation preserves the asphyxia label
Tanh distortion and room reverberation are applied to positive and negative samples; it is assumed these transformations do not alter the clinical condition represented by the cry.
domain assumption FedAvg on linear SVM weights converges to a global model comparable to a centrally trained model
Standard FL assumption [7]; the paper provides no convergence analysis and the claim of outperformance depends on this premise.
domain assumption The cleaned 10s mobile recording resembles the 1s training clips
Inference pipeline records 10s, extracts a segment via VAD, and classifies; the mismatch between 1s training clips and 10s recordings is not analyzed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HumekaFL: Automated Detection of Neonatal Asphyxia Using Federated Learning." pith.science (2026). https://pith.science/paper/KB5VAAHT

@misc{pith2026241201167,
  author       = {Pith},
  title        = {Pith review of: HumekaFL: Automated Detection of Neonatal Asphyxia Using Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KB5VAAHT}},
  note         = {Machine review of arXiv:2412.01167}
}

read the original abstract

Birth Apshyxia (BA) is a severe condition characterized by insufficient supply of oxygen to a newborn during the delivery. BA is one of the primary causes of neonatal death in the world. Although there has been a decline in neonatal deaths over the past two decades, the developing world, particularly sub-Saharan Africa, continues to experience the highest under-five (<5) mortality rates. While evidence-based methods are commonly used to detect BA in African healthcare settings, they can be subject to physician errors or delays in diagnosis, preventing timely interventions. Centralized Machine Learning (ML) methods demonstrated good performance in early detection of BA but require sensitive health data to leave their premises before training, which does not guarantee privacy and security. Healthcare institutions are therefore reluctant to adopt such solutions in Africa. To address this challenge, we suggest a federated learning (FL)-based software architecture, a distributed learning method that prioritizes privacy and security by design. We have developed a user-friendly and cost-effective mobile application embedding the FL pipeline for early detection of BA. Our Federated SVM model outperformed centralized SVM pipelines and Neural Networks (NN)-based methods in the existing literature

Figures

Figures reproduced from arXiv: 2412.01167 by the authors.

**Figure 1.** System architecture used to represent every 1000 ms cry clip from the Baby Chillanto database. The Baby Chillanto Infant Cry database is small in size. This can present a challenge for ML/DL pipelines that typically benefit from larger datasets to enhance learning speed and performance. Consequently, we augmented the data. In our data augmentation process, we chose to bootstrap the dataset using two techniques: tanh… view at source ↗

**Figure 2.** HumekaFL Diagnostic Process Federated SVM (FedSVM). To classify newborns’ cries in a federated setting, we first decide the ML model to employ in training our data on distributed clients. HumekaFL leverages Support Vector Machine (SVM) to train the Baby Chillanto Dataset on ten (10) virtual silos (hospitals). There are few important considerations justifying this choice: (1) the size of the training dataset, which… view at source ↗

**Figure 3.** HumekaFL mobile application’s screens (a) Training Loss vs Communication rounds (b) Average Train Accuracy vs Communication rounds 4 CONCLUSION In this work, we propose HumekaFL an FL-based mobile application for early detection of newborn birth asphyxia. Our SVM-based federated pipeline, FedSVM, outperformed existing centralized SVM and NN-based models using the Baby Chillanto dataset. Data augmentation, hyperp… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 4 canonical work pages

[1]

Cynthia Dwork. 2006. Differential privacy. InInternational colloquium on automata, languages, and programming. Springer, 1–12

2006
[2]

Google. 2011. Real-time communication for the web. Retrieved April 02, 2024 from https://webrtc.org/

work page 2011
[3]

Chunyan Ji, Thosini Bamunu Mudiyanselage, Yutong Gao, and Yi Pan. 2021. A review of infant cry analysis and classification. EURASIP Journal on Audio, Speech, and Music Processing 2021, 1 (2021), 8

work page 2021
[4]

Yasmina Kheddache and Chakib Tadj. 2019. Identification of diseases in newborns using advanced acoustic features of cry signals. Biomedical signal processing and control 50 (2019), 35–44

work page 2019
[5]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimiza- tion. arXiv preprint arXiv:1412.6980 (2014)

arXiv 2014
[6]

Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2 (2020), 429–450

2020
[7]

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics. PMLR, 1273–1282

2017
[8]

Charles C Onu, Innocent Udeogu, Eyenimi Ndiomu, Urbain Kengni, Doina Precup, Guilherme M Sant’Anna, Edward Alikor, and Peace Opara. 2017. Ubenwa: Cry- based diagnosis of birth asphyxia. arXiv preprint arXiv:1711.06405 (2017)

arXiv 2017

Show all 9 references

[9]

Orion Fausto Reyes-Galaviz, Sergio Daniel Cano-Ortiz, and Carlos Alberto Reyes- García. 2008. Evolutionary-neural system to classify infant cry units for patholo- gies identification in recently born babies. In 2008 Seventh Mexican international conference on artificial intell...

2008

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Cynthia Dwork. 2006. Differential privacy. InInternational colloquium on automata, languages, and programming. Springer, 1–12

2006

[2] [2]

Google. 2011. Real-time communication for the web. Retrieved April 02, 2024 from https://webrtc.org/

work page 2011

[3] [3]

Chunyan Ji, Thosini Bamunu Mudiyanselage, Yutong Gao, and Yi Pan. 2021. A review of infant cry analysis and classification. EURASIP Journal on Audio, Speech, and Music Processing 2021, 1 (2021), 8

work page 2021

[4] [4]

Yasmina Kheddache and Chakib Tadj. 2019. Identification of diseases in newborns using advanced acoustic features of cry signals. Biomedical signal processing and control 50 (2019), 35–44

work page 2019

[5] [5]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimiza- tion. arXiv preprint arXiv:1412.6980 (2014)

arXiv 2014

[6] [6]

Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2 (2020), 429–450

2020

[7] [7]

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics. PMLR, 1273–1282

2017

[8] [8]

Charles C Onu, Innocent Udeogu, Eyenimi Ndiomu, Urbain Kengni, Doina Precup, Guilherme M Sant’Anna, Edward Alikor, and Peace Opara. 2017. Ubenwa: Cry- based diagnosis of birth asphyxia. arXiv preprint arXiv:1711.06405 (2017)

arXiv 2017