{"id":"908557fc-0dac-46cc-b22a-ff570dce27fd","arxiv_id":"2506.15694","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A GA with supposedly parallel fitness evaluation tunes an MLP on three disease datasets and reports 99.12%, 94.87%, and 100% accuracy, but the test set is used as the tuning objective.","lead":"This paper combines kernel PCA, a multilayer perceptron, and a genetic algorithm with parallel fitness checks to predict disease from three public medical datasets. The reported accuracies look high, but the algorithm tunes on the test set and calls thread-based parallelism multiprocessing, so the headline numbers need scrutiny.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set fitness selection undercuts the headline accuracies; the reported best numbers are the optimized objective on the same split, and internal table inconsistencies make them unreliable.","rationale":"Reading in good faith, the paper is an engineering report whose contributions are a thread-parallel GA tuner, kernel PCA preprocessing, and a GUI for MLP-based disease prediction. The headline accuracies are the central evidence for both the tuning method and the comparison with other optimizers. The reader's identification of test-set fitness as the load-bearing weakness is exactly right: Table 2 makes it explicit that the GA fits to test accuracy, and Algorithm 1 tracks the best individual by that score. That makes the final numbers the optimized objective values, not unbiased estimates. This is a methodological flaw, not a question of style or competitive framing. The internal inconsistencies in Table 6 and the surrounding prose reinforce the problem: if the optimal configurations cannot be stated consistently, the reported accuracies are not reproducible from the manuscript alone. I agree with the reader that the evidentiary value of the headline accuracy numbers and the optimizer comparison collapses. My verdict recommendation is UNCHANGED because the reader already recommended REJECT and this stress test confirms that verdict; it does not introduce a new concern that would shift it further. I deliberately avoid speculating about author intent; the issue is visible in the manuscript's own stated protocol.","tokens_in":10080,"tokens_out":3608,"duration_ms":38749,"concrete_test":"Re-run the MIGA pipeline on each of the three datasets with a three-way split: 64% training, 16% validation, and 20% untouched test. Replace the Table 2 fitness function with validation accuracy instead of test accuracy. After the 10-generation search, retrain the best configuration on training plus validation and evaluate it once on the untouched test set. Apply the identical protocol to grid search, random search, and Bayesian optimization, and repeat over at least 10 random seeds to obtain error bars. If the independently evaluated MIGA-MLP test accuracy no longer reaches 99.12% / 94.87% / 100%, or no longer beats the comparators under the same protocol, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MIGA-tuned MLPs achieve 99.12% accuracy on breast cancer, 94.87% on Parkinson's, and 100% on CKD, and that these results beat grid search, random search, and Bayesian optimization. The load-bearing condition is that the reported accuracies are unbiased estimates on data not used for model selection. Table 2 states the GA fitness function is 'Accuracy score on test set', and Algorithm 1 in Section 3.5.1 explicitly tracks the best individual by that fitness. Consequently, the 'best' accuracy in Tables 3-5 is the maximum of the exact statistic that the GA is optimizing on the exact split later reported; it is a selection artifact, not an independent measurement. The comparison against other optimizers is also uninterpretable unless every method used the same held-out protocol, which the paper does not document. The unreliability is compounded by internal inconsistencies: Table 6 reports 95.00% for Parkinson's while Table 3's best is 0.9487; the breast-cancer text says 50 hidden units while Table 6 says 100; and the CKD text says 50 hidden units while Table 6 says 100. These discrepancies mean the precise configurations behind the headline numbers are not uniquely specified. No code, reproducibility package, or external validation is provided, so the claimed accuracies and the claimed superiority over other optimizers cannot be credited without an independent holdout evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents MIGA, a genetic algorithm with parallel fitness evaluation for tuning an MLP classifier, combined with RBF kernel PCA for dimensionality reduction. The framework is evaluated on three public medical datasets (Wisconsin Diagnostic Breast Cancer, Parkinson's Telemonitoring, and chronic kidney disease), reporting best accuracies of 99.12%, 94.87%, and 100%, a tuning-time reduction of about 60%, and a GUI for clinicians. The paper claims these results outperform grid search, random search, and Bayesian optimization.","tokens_in":10321,"tokens_out":6244,"duration_ms":63559,"significance":"The proposed pipeline is clearly described and the use of public datasets makes the application concrete; a low-code GUI for hyperparameter tuning would be practically useful if the underlying claims held. However, the central quantitative claims are supported only by a circular evaluation protocol and contain internal inconsistencies, and no code or reproducibility package is provided. As a result, the current evidence does not establish the claimed superiority or speedup.","major_comments":[{"comment":"The GA fitness function is defined as accuracy on the test set (Table 2), and Algorithm 1 uses that same fitness to track the best individual. The best accuracies in Tables 3–5 are therefore the maximum of the test-set accuracy over 10 generations, i.e., the optimized objective on the reported split. This makes the headline accuracies selection artifacts, not independent estimates of generalization. Please re-run with fitness evaluated on a validation split and report final performance on a held-out test set (or use nested cross-validation), and ensure all compared optimizers use the same protocol.","section":"Table 2; Algorithm 1, Section 3.5.1"},{"comment":"The optimal configurations reported in the text do not match Table 6: the text gives 50 hidden units for breast cancer while Table 6 gives 100; the text gives 50 hidden units for CKD while Table 6 gives 100; and the Parkinson's accuracy in Table 6 is 95.00% although Table 3 reports a best of 0.9487. These inconsistencies mean the exact configuration behind each headline accuracy is not uniquely specified, so the results are not reproducible as written.","section":"Section 4, text near Table 6 and Table 6"},{"comment":"The claimed 59–61% time reduction relies on 'multithreading' via ThreadPoolExecutor, which under CPython does not provide true parallelism for CPU-bound Python-level MLP fitting because of the GIL. No profiling, hardware details, per-generation timings, or repeated runs are provided, so the speedup is not supported. The authors should either use process-based parallelism (ProcessPoolExecutor) or demonstrate with profiling and core-utilization data that the threaded execution actually parallelizes the workload.","section":"Section 3.5; Table 2; Table 7 (timing)"},{"comment":"The abstract claims that MIGA outperforms grid search, random search, and Bayesian optimization, but no controlled comparison with these methods under identical splits appears in the manuscript. The second Table 7 lists accuracies from unrelated studies that use different datasets, preprocessing, and evaluation protocols. A direct comparison on the same train/test splits with identical metrics is required to support the superiority claim.","section":"Abstract; Section 4, second Table 7"}],"minor_comments":[{"comment":"Reference [20] for the Parkinson's dataset points to the UCI chronic kidney disease page rather than the Parkinson's Telemonitoring dataset.","section":"Declarations / Reference [20]"},{"comment":"Table 7 is used twice (timing logs and comparison with other models); the second table should be renumbered.","section":"Section 4"},{"comment":"The algorithm is called 'Multi-Processing' but the implementation described in the text and Table 2 uses multithreading; the terminology should be reconciled.","section":"Section 3.5"},{"comment":"The Parkinson's Telemonitoring dataset is originally a regression dataset; please state how the binary status target was derived and report class balance, since this affects accuracy interpretation.","section":"Section 3.1.2"},{"comment":"There are frequent typos, including 'Parkison'/'Parkision' for Parkinson, 'stage of the art' for state of the art, and 'treating breast cancer' in a figure caption; a careful proofread is needed.","section":"Throughout"},{"comment":"Please clarify whether kernel PCA was fitted on the training split only; if fitted on the entire dataset before splitting, the reported results would be optimistically biased.","section":"Section 3.2/3.3"}],"recommendation":"reject","confidential_remarks":"The paper's core claims rest on a circular evaluation and cannot be verified without code; the inconsistencies in the result tables compound the problem. The contribution may be salvageable as a software/tool paper after a complete re-evaluation, but as submitted it does not meet the bar. I would not ask for a major revision because the required changes involve replacing the central experimental evidence, not local corrections."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a standard GA-tunes-MLP paper with a circular evaluation that kills the headline accuracies, and the speedup claim is shaky. But it is clearly written and honest about what it did, so it's not a scam—just not a publishable result as it stands.\n\nWhat's new: basically a ThreadPoolExecutor around the fitness evaluation, plus a GUI. That's a well-known parallel-GA pattern, and the paper's own references show prior GA-MLP tuning. The integration with KPCA and the GUI is a practical touch for a practitioner, but there's no new algorithm or theory.\n\nThe big problem: Table 2 says the fitness function is 'Accuracy score on test set.' So the GA is selecting hyperparameters that maximize accuracy on the exact same 20% held-out samples that are then reported as the result. That is structurally circular—the reported 99.12%, 94.87%, and 100% are the optimized objective, not independent measurements. There's no separate validation set, no cross-validation, no error bars. The comparison against grid/random/Bayesian is uninterpretable unless those methods used the same flawed protocol, and that is not documented.\n\nThe speedup claim has its own issues. ThreadPoolExecutor does not reliably parallelize CPU-bound MLP fitting under CPython's GIL, and no profiling is shown. Worse, the numbers in Table 7 don't match the formula: they report a 59.0% reduction for breast cancer but (107.05-48.05)/107.05 is about 55.1%; for CKD, 60/71.46 is about 84.0%, not 60.0%. So the timing table is internally inconsistent.\n\nThere are also configuration inconsistencies: the breast cancer text says 50 hidden units, Table 6 says 100; same for CKD. Table 6 gives Parkinson's accuracy as 95.00%, but Table 3's best is 0.9487 (94.87%). And the comparison table lists the proposed method at 99.00% for breast cancer, not the 99.12% from the abstract.\n\nWhat the paper does well: the writing is clear, the algorithm is described in enough detail to re-implement, and the GUI is a nice practical addition. The authors don't hide the test-set fitness—it's right there in Table 2. That is transparent, but transparency doesn't fix the flaw.\n\nBottom line: this is a decent undergraduate project or workshop demo, but it doesn't meet the bar for a peer-reviewed article. I would not cite it for the results, and I would not send it to external referees—the circularity and inconsistencies are visible in the manuscript, so a desk reject is appropriate. There's no point burning reviewer time on an evaluation that would have to be redone from scratch.","headline":"Routine GA-MLP tuning paper with a circular test-set fitness that invalidates its headline accuracies and a speedup table that doesn't add up.","tokens_in":10895,"tokens_out":3464,"would_cite":false,"duration_ms":31035,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multiprocessing genetic algorithm tunes a multilayer perceptron to 99.12% accuracy on breast cancer, 94.87% on Parkinson's, and 100% on chronic kidney disease, and cuts tuning time by about 60%.","keywords":["multilayer perceptron","multiprocessing genetic algorithm","hyperparameter optimization","kernel principal component analysis","parallel processing","breast cancer","Parkinson's disease","chronic kidney disease"],"falsifier":"Re-run the MIGA pipeline on all three datasets with the test set locked away until after tuning: optimize with a separate validation split as fitness, then measure accuracy once on the untouched test set. If held-out accuracy falls materially below 99.12% (breast cancer), 94.87% (Parkinson's), and 100% (CKD), the reported numbers were an artifact of selecting on the test set. A cheaper check is to repeat the 80/20 split with different random seeds and record how much the MIGA-selected configuration's accuracy varies across folds.","tokens_in":9815,"feed_emoji":"🧬","tokens_out":12071,"duration_ms":108791,"temperature":0.7,"pith_summary":"This paper claims that a disease-prediction pipeline can be made both more accurate and faster to tune by combining kernel PCA for nonlinear feature reduction with a multilayer perceptron whose hyperparameters are chosen by a modified multiprocessing genetic algorithm (MIGA). The MIGA evaluates candidate hyperparameter configurations in parallel across CPU threads, which the authors report cuts tuning time by roughly 60% relative to a single-threaded genetic algorithm. On three standard medical datasets the tuned MLP reaches 99.12% accuracy on breast cancer, 94.87% on Parkinson's disease, and 100% on chronic kidney disease, outperforming grid search, random search, and Bayesian optimization in the paper's comparisons. If true, this would give clinicians and researchers a largely automatic, much cheaper route to high-accuracy tabular disease classifiers.","feed_headline":"Parallel tuning cuts MLP search time ~60%, hits 99%+ accuracy","feed_subtitle":"MLP tuned by a parallel genetic algorithm scores 99.12% on breast cancer, 94.87% on Parkinson's, 100% on kidney disease.","key_machinery":"The load-bearing mechanism is the multiprocessing interface genetic algorithm (MIGA), a genetic algorithm whose fitness-evaluation step—training and scoring an MLPClassifier on the test set for each candidate hyperparameter configuration—is executed in parallel across multiple threads instead of sequentially. Each chromosome is the tuple [hidden-layer size, activation function, learning rate, solver]; fitness is the accuracy score on the test set; the top 50% of the population is kept by elitist selection, recombined by uniform crossover, and mutated with probability 0.1. The parallel evaluation is what the paper credits for the roughly 60% reduction in tuning time, while the preceding kernel PCA stage supplies the nonlinear features that the authors say improve classification relative to a standard GA.","core_discovery":"The paper's central claim is that its proposed framework—kernel PCA with a radial basis function kernel keeping 95% of the variance, a multilayer perceptron classifier, and a multiprocessing genetic algorithm (MIGA) that parallelizes hyperparameter search—delivers high disease-classification accuracy at sharply reduced tuning cost. Each candidate MLP configuration is encoded as a chromosome (hidden-layer size, activation, learning rate, solver), and a population of 10 is evolved over 10 generations with elitist selection, uniform crossover, and a 10% mutation rate, while fitness evaluations run in parallel through a thread pool. The reported best accuracies are 99.12% for the Wisconsin Diagnostic Breast Cancer dataset, 94.87% for the Parkinson's telemonitoring dataset, and 100% for the chronic kidney disease dataset, with tuning time reduced by approximately 60% compared with a standard single-threaded GA. The per-generation logs show the best configuration typically appearing within the first two generations and stabilizing afterward, which the authors present as evidence of efficient convergence.","pith_inferences":["Not explored by the authors but a direct consequence: the MIGA wrapper only needs a fitness function, so the same parallel-tuning interface should transfer to SVM, logistic regression, or gradient-boosted tree classifiers on the same datasets, where per-evaluation cost is lower than an MLP's.","The reported ~60% speedup is measured against a single-threaded GA; a matched-budget comparison against parallelized Bayesian optimization or random search with early stopping would be a fairer practical yardstick and is not reported.","Because the best configuration emerges by generation 2, the time saved by parallelization could be reinvested in larger populations or more generations to search harder datasets—an option the paper leaves implicit.","The paper's own tables do not fully agree—breast-cancer accuracy appears as both 99.12% and 99.00%, and the CKD optimal configuration differs between the text and Table 6—so a reader benchmarking this framework should reconcile the reported figures first."],"forward_implications":["Evolutionary hyperparameter search becomes practical in compute- or time-constrained settings: the same search quality costs about 40% of the wall-clock time, so facilities with limited hardware could run GA tuning where it was previously too slow.","The bundled graphical interface means the full pipeline—loading a CSV, choosing the target, reducing dimensions, tuning, and predicting—can be operated without writing code, which is the paper's stated route to clinical adoption.","The per-generation tables indicate the best configuration appears by generation 2 on these datasets, so a short evolutionary search captures most of the accuracy gain and longer runs mainly confirm the plateau.","Kernel PCA contributes to the result as well as the GA: the paper reports that the nonlinear reduction improved classification where a standard GA without it did not, so both components—feature extraction and optimizer—carry the performance.","On all three datasets the MIGA-tuned MLP matches or beats the cited published baselines, including Bayesian-optimized SVM, PSO-tuned CNN, and GA-tuned neural networks, which is the paper's evidence that the framework is competitive with established tuning methods."],"supporting_citations":[{"why":"The authors' own earlier MLP with kernel PCA for chronic kidney disease, which reached 100% accuracy and is the direct ancestor of this framework's feature-extraction step.","marker":"[3]"},{"why":"The genetic-algorithm source cited for selection, crossover, and mutation mechanics that the MIGA modifies, and referenced again in the discussion for the parallelization gains.","marker":"[7]"},{"why":"The real-coded GA for MLP hyperparameter tuning (MLPRGA+5) whose method this work extends and whose 96% breast-cancer accuracy is a comparison baseline.","marker":"[12]"},{"why":"The GA-MLP hybrid with PCA for chronic kidney disease (98.54%) that the proposed 100% CKD result is compared against.","marker":"[13]"},{"why":"The genetically optimized neural network for Parkinson's disease (95%) that serves as the nearest baseline for the reported 94.87%.","marker":"[19]"},{"why":"The PSO-tuned CNN for mammography (98.23% DDSM, 97.98% MIAS) used as a breast-cancer comparison baseline.","marker":"[22]"},{"why":"Cited for the claim that genetic algorithms are inherently parallelizable, which motivates the multiprocessing design.","marker":"[26]"},{"why":"The Bayesian-optimized SVM for Parkinson's (92.30%) used as the Bayesian-optimization comparison baseline.","marker":"[29]"}],"fun_headline_variants":["Parallel GA tunes MLP 60% faster, hits up to 100% accuracy","MIGA parallel GA: 100% CKD, 99% breast, 95% Parkinson's","Parallel GA cuts MLP tuning time ~60%, boosts accuracy","MLP hyperparameter search 60% faster with parallel GA","Parallel genetic algorithm tunes MLP, hits 99-100% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that using test-set accuracy as the genetic algorithm's fitness function—stated in Table 2—still gives an unbiased estimate of final performance, since the same test split both steers the search and is later reported as the result.","fun_headline_variants_meta":{"raw":{"variants":["Parallel GA tunes MLP 60% faster, hits up to 100% accuracy","MIGA parallel GA: 100% CKD, 99% breast, 95% Parkinson's","Parallel GA cuts MLP tuning time ~60%, boosts accuracy","MLP hyperparameter search 60% faster with parallel GA","Parallel genetic algorithm tunes MLP, hits 99-100% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000567,"raw_usage":{"total_tokens":2730,"prompt_tokens":1035,"completion_tokens":1695,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":1607}},"tokens_in":651,"tokens_out":1695,"duration_ms":11858,"temperature":1.0,"reasoning_tokens":1607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:25:15.545734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the MIGA pipeline on all three datasets with the test set locked away until after tuning: optimize with a separate validation split as fitness, then measure accuracy once on the untouched test set. If held-out accuracy falls materially below 99.12% (breast cancer), 94.87% (Parkinson's), and 100% (CKD), the reported numbers were an artifact of selecting on the test set. A cheaper check is to repeat the 80/20 split with different random seeds and record how much the MIGA-selected configuration's accuracy varies across folds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The authors' own earlier MLP with kernel PCA for chronic kidney disease, which reached 100% accuracy and is the direct ancestor of this framework's feature-extraction step."},{"cited_title":"Genetic algorithm to optimize the SVM and K -means algorithms for mapping of mineral prospectivity,","cited_arxiv_id":null,"evidence_quote":"The genetic-algorithm source cited for selection, crossover, and mutation mechanics that the MIGA modifies, and referenced again in the discussion for the parallelization gains."},{"cited_title":"Chronic kidney disease prediction using boosting techniques based on clinical parameters,","cited_arxiv_id":null,"evidence_quote":"The real-coded GA for MLP hyperparameter tuning (MLPRGA+5) whose method this work extends and whose 96% breast-cancer accuracy is a comparison baseline."},{"cited_title":"Explainable Machine Learning Model for Chronic Kidney Disease Prediction,","cited_arxiv_id":null,"evidence_quote":"The GA-MLP hybrid with PCA for chronic kidney disease (98.54%) that the proposed 100% CKD result is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The genetically optimized neural network for Parkinson's disease (95%) that serves as the nearest baseline for the reported 94.87%."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The PSO-tuned CNN for mammography (98.23% DDSM, 97.98% MIAS) used as a breast-cancer comparison baseline."},{"cited_title":"Recent trends in prediction of chronic kidney disease using different learning approaches : a systematic literature review,","cited_arxiv_id":null,"evidence_quote":"Cited for the claim that genetic algorithms are inherently parallelizable, which motivates the multiprocessing design."},{"cited_title":"Enhancing the Early Detection of Chronic Kidney Disease : A Robust Machine Learning Model,","cited_arxiv_id":null,"evidence_quote":"The Bayesian-optimized SVM for Parkinson's (92.30%) used as the Bayesian-optimization comparison baseline."}],"review_version":1}