{"id":"b18c7e2b-1c67-48d5-9b27-a6a6d69263ed","arxiv_id":"2502.15762","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"SmartEdge integrates edge and cloud computing with ensemble voting to predict diabetes, reporting accuracy around 81% on the PIMA dataset.","lead":"The paper presents SmartEdge, an edge and cloud computing system for diabetes prediction using ensemble machine learning, and reports that voting models improve accuracy by about 5% over single models. It is a systems integration paper that combines known ML and edge computing into an end-to-end deployment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 5% ensemble accuracy gain is not visible in Table II (80.9% vs 77.8% best single = 3.0 pp) and rests on one 70/10/20 split of 537 records, so the headline improvement could be split variance.","rationale":"The reader's weakest-assumption analysis correctly identifies the single 70/10/20 split as the load-bearing weakness. My stress-test finds the same concern and adds that even the reported point estimates fail to show a 5% gain: 0.8086 vs 0.7784 is about 3 percentage points, and the second voting ensemble is essentially tied with the best single model. The system implementation itself is plausible: it uses standard components (scikit-learn, FogBus2, PIMA data), and the latency results are presented as descriptive measurements, so I do not object to the architecture per se. The problem is exclusively that the headline accuracy claim is an unsupported point estimate from one split with no variance or significance testing. A repeated cross-validation with confidence intervals and a paired significance test is the natural, decisive check. Since the reader already rejected the paper for this core evidential gap, my verdict is unchanged.","tokens_in":10272,"tokens_out":3341,"duration_ms":34667,"concrete_test":"Re-run the same preprocessing on the 537-record PIMA dataset with repeated stratified cross-validation (e.g., 10 repeats of 5-fold), computing mean and 95% CI for LR, RF, GBoost, DT, and both voting ensembles, plus McNemar's test comparing the best ensemble to the best single model. Also recompute the Table II comparison to state both percentage-point and relative gains. If the mean ensemble gain over the best single model is not at least 5 percentage points (or the stated interpretation) with non-overlapping CI and McNemar p < 0.05, the central accuracy claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim (Abstract; Section VI-D) is that ensemble voting improves diabetes-prediction accuracy by 5% relative to a single model. The evidence offered is Table II, from one 70/10/20 split of 537 preprocessed PIMA records. Two problems make this claim load-bearing. First, the arithmetic does not literally support 5%: the best voting ensemble (Voting SVM, DT, LR) reaches 0.8086 while the best reported single model (LogisticRegression) reaches 0.7784, a gain of 3.02 percentage points (about 3.9% relative). The second voting row (Voting RF SVM LR, 0.784) is within noise of the best single model. Second, with 537 records and a 20% test split (~107 test cases), a 3-percentage-point difference is a handful of correct predictions; no standard deviation, confidence interval, cross-validation, or repeated runs are reported, and the paper even says 'we first make a run' (Section VI-C). The claim 'improves consistently accuracy by 5%' is therefore not established. For the claimed 5% gain to hold, the ensemble would need to beat every component (or at least the best single model) by approximately 5 pp with variance small enough that the gain cannot be explained by split luck. The manuscript does not provide that evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SmartEdge, an end-to-end edge-cloud system for diabetes prediction built on the FogBus2 framework. The authors deploy single and ensemble machine-learning models on Raspberry Pi edge nodes and cloud VMs, evaluate prediction accuracy on the PIMA Indian Diabetes dataset, and report latency and response-time measurements. The central claim, stated in the Abstract and Sections VI-D and VI-F, is that ensemble voting improves prediction accuracy by 5% over single-model prediction, and that the edge-cloud architecture provides low-latency responses.","tokens_in":10567,"tokens_out":5171,"duration_ms":51057,"significance":"If the claims were fully supported, SmartEdge would be a useful engineering demonstration of deploying ensemble learning in an edge-cloud IoMT pipeline. The use of a public dataset, a deployable framework, and a comparison of several model configurations are commendable aspects. However, the reported evidence does not support the 5% accuracy improvement claim, and the evaluation lacks the statistical rigor needed to establish the stated benefits. With corrected analysis and repeated cross-validation, the system could still represent a modest but valid contribution to edge-based healthcare machine-learning deployment.","major_comments":[{"comment":"The stated claim that ensemble voting improves accuracy by 5% is not supported by Table II. The best ensemble (Voting SVM, DT, LR) achieves 0.8086, while the best single model (LogisticRegression) achieves 0.7784, which is an improvement of 3.02 percentage points (approximately 3.9% relative), not 5%. The second ensemble (Voting RF SVM LR) achieves 0.784, within 0.6 pp of the best single model. The sentence 'improves consistently accuracy by 5%' in Section VI-D (and repeated in VI-F) therefore overstates the reported results.","section":"Abstract; Section VI-D; Table II"},{"comment":"The evaluation uses a single 70/10/20 split of 537 records. With a 20% test set (about 107 samples), the difference between 0.8086 and 0.7784 corresponds to roughly three additional correct predictions, which is well within sampling variability. No cross-validation, repeated runs, bootstrap confidence intervals, or significance tests are reported. Consequently, the claim that ensemble voting 'consistently' outperforms single models is not statistically supported. A repeated stratified k-fold or repeated hold-out with confidence intervals is needed before drawing conclusions.","section":"Section VI-C and VI-D"},{"comment":"The composition of the voting ensemble is inconsistent across the paper. Section V states that the voting scheme utilized RF, CatBoost, and LR, while Section VI-C lists SVM, RF, Logistic Regression, and Decision Tree. Table II reports two voting configurations, Voting SVM/DT/LR and Voting RF/SVM/LR, and no CatBoost results appear anywhere. This inconsistency prevents reproduction of the reported experiments.","section":"Section V and Table II"},{"comment":"The latency and response-time results are presented only as figures (Figures 7-13) without numeric summary statistics, variance, or statistical comparisons. The text claims that edge nodes achieve lower latency for user-to-master communication compared to cloud, but the magnitude and variability of these differences are not reported. Since low latency is a core contribution of the SmartEdge architecture, these measurements need to be quantified and analyzed.","section":"Sections VI-E and VI-F"}],"minor_comments":[{"comment":"The conclusion states that SmartEdge uses deep learning and has been applied to heart disease analysis, but neither deep learning nor heart disease experiments appear in the paper; the conclusion should be aligned with the actual experimental content.","section":"Section VII"},{"comment":"The reference '[91]' in the data scaling sentence is out of range; the reference list has 22 entries.","section":"Section IV.A"},{"comment":"The phrase 'Logical Regression' should be 'Logistic Regression', and 'Fogbus2' is spelled inconsistently as 'FogBus2'.","section":"Section VI.C"},{"comment":"Several figures lack clear axis labels and legible legends; for example, Figure 6 has no y-axis title and Figures 7-9 show time series with unexplained units.","section":"Figures 7-13"},{"comment":"The checkmark symbols in Table I are ambiguous; a legend or explanatory text would help readers interpret which features each work supports.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central quantitative claim is contradicted by its own table, and the evaluation design is insufficient to establish the claimed improvement. These issues are fixable with additional experiments and careful rewriting, but the current version is not ready for publication. I recommend major revision rather than rejection because the underlying system and deployment concept appear sound and the authors could produce a defensible empirical evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nSmartEdge is a straightforward engineering case study: the authors bolt ensemble voting onto FogBus2 to run diabetes prediction on Raspberry Pi edge nodes and cloud VMs, and measure latency, response time, and accuracy. The systems part is concrete and the paper is honest about its setup. That's the good news.\n\nThe bad news is the headline. The abstract and Section VI-D claim ensemble voting 'improves consistently accuracy by 5%' versus a single model. Their own Table II shows best ensemble at 0.8086 and best single at 0.7784, a gain of 3.02 percentage points (about 3.9% relative). A second voting combo at 0.784 is within noise. With a single 70/10/20 split of 537 records (about 107 test cases), that 3-point gap is a handful of predictions. No cross-validation, no repeated runs, no confidence intervals. The paper even says 'we first make a run,' which sounds like one shot. So the central quantitative claim is not established.\n\nThe novelty is modest: the related work table shows several earlier systems using ensembles and edge/cloud in diabetes prediction. The claimed gap—'full scale' edge/fog/cloud—is a marginal distinction. The ML pipeline is scikit-learn standard.\n\nWhere I'd give credit: the end-to-end deployment across four distributed scenarios is real work, and the latency/response-time measurements are useful for practitioners. The related work table is a decent map. The discussion of deployment trade-offs is sensible.\n\nSoft spots beyond the 5% claim: the conclusion overstates, mentioning deep learning and heart disease without experimental support. That's an overstatement, though not central.\n\nNet: the paper is a marginal application paper with a load-bearing but fixable flaw. If the authors correct the arithmetic, add cross-validation or error bars, and temper the claims, it could be a usable systems paper. As submitted, the central assertion is unsupported.\n\nI'd send it to review—it deserves referee time for the systems evaluation—but expect major revision. I wouldn't cite it until the numbers hold up.","headline":"The 5% ensemble accuracy claim doesn't survive contact with the paper's own Table II, but the edge-cloud deployment is a real case study worth considering under heavy revision.","tokens_in":11090,"tokens_out":2419,"would_cite":false,"duration_ms":23301,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Ensemble voting improves diabetes prediction accuracy by 5% over single models in an integrated edge-cloud system.","keywords":["edge computing","cloud computing","ensemble learning","voting classifier","diabetes prediction","internet of medical things","FogBus2"],"falsifier":"Run the same preprocessing and models on the same 537 records across ten different random 70/10/20 splits; if the ensemble's accuracy advantage over the best single model is not positive in most splits, the claimed 5% improvement does not hold.","tokens_in":10087,"feed_emoji":"🩺","tokens_out":5281,"duration_ms":50580,"temperature":0.7,"pith_summary":"SmartEdge is an end-to-end system that places diabetes-prediction machine learning on edge nodes near patients, with cloud servers as a fallback for heavier tasks. The paper's central claim is that combining Random Forest, Logistic Regression, and Support Vector Machines through a voting ensemble improves prediction accuracy by about 5% over any single model, reaching roughly 80.9% accuracy on the Pima Indian Diabetes dataset. The authors argue this matters because IoMT patient monitoring is latency-sensitive, and pushing prediction to the edge avoids the delays of sending every request to the cloud. The contribution is a demonstration that ensemble voting can be embedded in a working edge-cloud healthcare pipeline without losing the latency benefit.","feed_headline":"Ensemble voting lifts diabetes prediction accuracy by 5%","feed_subtitle":"SmartEdge runs the vote across edge and cloud nodes, reaching about 81% accuracy on the PIMA dataset.","key_machinery":"The central mechanism is a voting ensemble of three classifiers. In hard voting, each classifier outputs a class and the majority wins; in soft voting, each classifier outputs class probabilities and the combined probabilities decide. SmartEdge deploys these voters as tasks on edge worker nodes managed by a broker node, which selects the least-loaded worker, with cloud servers available for latency-tolerant work. The argument turns on distributing model execution to edge nodes so prediction happens close to the patient, while the cloud remains a fallback for compute-intensive tasks.","core_discovery":"On the Pima Indian Diabetes dataset, after removing records with missing values so that 537 records remain, the voting ensemble of SVM, Decision Tree, and Logistic Regression reaches an accuracy of 0.8086, compared with 0.7784 for logistic regression, 0.7722 for random forest, 0.7667 for gradient boosting, and 0.7037 for decision tree. The paper states this as a consistent 5% accuracy improvement from ensemble voting. The same voting approach holds when the voters are distributed across Raspberry Pi edge nodes, a single Raspberry Pi, or cloud virtual machines, with edge-path latencies measured in tens of milliseconds and cloud paths slower. The authors position SmartEdge as an end-to-end integrated edge-cloud healthcare system rather than a standalone model comparison.","pith_inferences":["Beyond the paper, the 5% accuracy gain is estimated from a single 70/10/20 split of 537 records; repeated splits or cross-validation could show that the gap varies substantially.","Because the ensemble executes three models per request, the accuracy gain trades directly against edge-node resource use, which is why the paper itself suggests disabling ensemble voting under bandwidth constraints.","A testable extension is to run the same voting ensemble on a larger, more recent diabetes cohort and report the spread of accuracy differences to see whether the advantage persists outside Pima.","The same edge-cloud voting architecture could generalize to other IoMT prediction tasks that share the same risk-factor input pattern, such as hypertension or cardiac risk screening."],"forward_implications":["A healthcare deployment can obtain about 81% diabetes-prediction accuracy on Pima-style risk factors without sending every prediction to the cloud.","Latency-critical monitoring can run entirely on local edge nodes, while the cloud path is recommended only for heavy, latency-tolerant tasks.","Ensemble voting consistently beats the best single classifier in these experiments, regardless of whether the voters run on one device, multiple edge nodes, or cloud virtual machines.","Deploying more edge nodes increases broker response time slightly, because the master must check each worker's load before assigning a task."],"supporting_citations":[{"why":"Supplies FogBus2, the backend framework that distributes prediction tasks across edge and cloud nodes.","marker":"[5]"},{"why":"Motivates the voting-classification ensemble approach for diabetes prediction.","marker":"[13]"},{"why":"Provides the Pima Indian diabetes classification baseline using Random Forest, one of the compared algorithms.","marker":"[7]"},{"why":"Provides the Python implementations of the classifiers and voting schemes used in the experiments.","marker":"[21]"},{"why":"Is the Pima Indian Diabetes dataset that all accuracy and latency measurements are based on.","marker":"[22]"}],"fun_headline_variants":["Ensemble voting on edge and cloud lifts diabetes accuracy by 5%","SmartEdge uses edge-cloud voting to gain 5% in diabetes prediction","Diabetes prediction improved 5% with edge-cloud ensemble voting","Edge and cloud ensemble voting boosts diabetes prediction by 5%","Voting ensemble across edge and cloud nodes ups diabetes accuracy 5%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single 70/10/20 split of the 537-record Pima Indian dataset is enough to measure the accuracy difference between ensemble and single models, since no repeated splits or cross-validation are reported.","fun_headline_variants_meta":{"raw":{"variants":["Ensemble voting on edge and cloud lifts diabetes accuracy by 5%","SmartEdge uses edge-cloud voting to gain 5% in diabetes prediction","Diabetes prediction improved 5% with edge-cloud ensemble voting","Edge and cloud ensemble voting boosts diabetes prediction by 5%","Voting ensemble across edge and cloud nodes ups diabetes accuracy 5%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000917,"raw_usage":{"total_tokens":3918,"prompt_tokens":913,"completion_tokens":3005,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":2914}},"tokens_in":529,"tokens_out":3005,"duration_ms":21964,"temperature":1.0,"reasoning_tokens":2914,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T19:55:10.241525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same preprocessing and models on the same 537 records across ten different random 70/10/20 splits; if the ensemble's accuracy advantage over the best single model is not positive in most splits, the claimed 5% improvement does not hold.","supporting_citations":[{"cited_title":"Fogbus2: a lightweight and distributed container -based framework for integration of iot - enabled systems with edge and cloud computing,","cited_arxiv_id":null,"evidence_quote":"Supplies FogBus2, the backend framework that distributes prediction tasks across edge and cloud nodes."},{"cited_title":"Voting Classification -Based Diabetes Mellitus Prediction Using Hypertuned Machine -Learning Techniques,","cited_arxiv_id":null,"evidence_quote":"Motivates the voting-classification ensemble approach for diabetes prediction."},{"cited_title":"Scikit-learn: Machine learning in Python,","cited_arxiv_id":null,"evidence_quote":"Provides the Python implementations of the classifiers and voting schemes used in the experiments."},{"cited_title":"Using the ADAP learning algorithm to forecast the onset of diabetes mellitus.,","cited_arxiv_id":null,"evidence_quote":"Is the Pima Indian Diabetes dataset that all accuracy and latency measurements are based on."}],"review_version":1}