{"id":"81c76c98-e5f0-464b-b0ae-e9ba13f7e451","arxiv_id":"2411.08195","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"On 862 dental X-rays, ensemble XGB and random forest models estimate age group (F1 73.26%) and gender (F1 77.53%) from tooth crown and pulp measurements.","lead":"This paper trains seven machine learning models on dental measurements from 862 bitewing radiographs to estimate age group and gender, reporting best F1 scores of 73.26% for age and 77.53% for gender. A generalist might read it to see how far off-the-shelf machine learning plus SHAP explanations can get in forensic dentistry, and where data-sharing and validation gaps still sit.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported F1 scores hinge on SMOTE being applied before the 80/20 split; if it was, the age and gender accuracies are optimistically biased and the central claim is unsubstantiated.","rationale":"Agreeing with the reader that the specific F1 numbers are the load-bearing claim. The reader's weakest assumption focused on manual measurement reliability; that is a valid external-validity concern but does not by itself invalidate the reported in-dataset performance. The more decisive internal threat is the SMOTE/split order. The paper's own diagram and section ordering suggest oversampling is done in pre-processing, before the dataset is split into train/test. SMOTE creates synthetic samples based on nearest neighbors in the full feature space; if it is applied before the split, the training and test sets are no longer independent, and standard evaluation metrics are optimistically biased. Five-fold cross-validation does not fix this if SMOTE is applied before the folds are created; it leaks within each fold as well. Because no code or data are available, this cannot be ruled out from the text, and the authors should be required to clarify and to provide a re-run with SMOTE placed inside the cross-validation loop. The abstract/Table 6 gender-F1 inconsistency is a second issue that needs a one-line correction but is less fundamental; it does not change the main recommendation. The reader's call for CONDITIONAL is appropriate: the paper should be published only after these checks are resolved.","tokens_in":12989,"tokens_out":4443,"duration_ms":45576,"concrete_test":"Obtain the authors' code or re-run the pipeline with one change: apply SMOTE only inside the training folds after splitting (e.g., via an imblearn Pipeline inside cross_val_score), keeping all other settings identical. If the mean F1 for age group and gender drops by more than about 2 points compared with the reported values, SMOTE-before-split leakage is confirmed and the headline numbers should be re-evaluated; if it does not drop, the order should still be documented explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated contribution is the specific F1 values: age group 73.26% (XGB) and gender 77.53% (RF/ensemble, Table 6). For these numbers to be valid generalization estimates, all target-dependent preprocessing must be confined to training data. The pipeline in Fig. 1 places \"Data Balancing (SMOTE)\" in the pre-processing stage, before the machine-learning stage, and Sec. 3.3 indicates that the 80/20 split is performed later, with 5-fold CV used for evaluation. If SMOTE is fit on the full dataset before splitting, synthetic minority-class samples derived from the whole dataset contaminate the training set, and the test set contains originals that were neighbors of those synthetic samples; the classifier is effectively evaluated on data it has already seen through the synthetic generation process. This directly inflates F1 and makes the reported numbers non-generalizable. No code or dataset is released, so the order cannot be independently checked. A related but secondary inconsistency is that the abstract credits the 77.53% gender F1 to Random Forest, whereas Table 6 lists that value as the \"Ensemble of above models\" and RF at 76.54%; this must be clarified. If leakage is present, the \"significant advancement\" claim collapses even though the measurement-reliability question raised in the reader's report is also valid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops machine learning classifiers for age-group and gender estimation from dental biometric measurements (coronal height, coronal pulp cavity height, and tooth coronal index) extracted from bitewing radiographs of 862 living individuals. The authors compare seven boosting/tree-based classifiers, propose an ensemble that trains per-tooth models and combines them via majority voting or non-trainable combiners, and use SHAP for interpretability. The headline results are an F1 score of 73.26% for age-group classification (XGB) and 77.53% for gender classification (attributed in the abstract to Random Forest, but to the ensemble in Table 6). The paper claims this constitutes a significant advancement in non-invasive dental forensic methods.","tokens_in":13356,"tokens_out":3978,"duration_ms":40921,"significance":"If the reported performance is a valid generalization estimate, the paper demonstrates a modestly useful, non-invasive screening tool for age and gender estimation that could complement expert analysis in forensic dentistry. The manuscript is careful in describing ethical approval, inclusion/exclusion criteria, and the clinical measurement protocol. A strength is the systematic comparison of several standard classifiers and the exploration of ensemble and non-trainable combiner strategies, which provides a useful baseline for the community. However, the central numerical claims are compromised by the apparent order of SMOTE before the train/test split, the lack of uncertainty quantification, and the inconsistency between the abstract and the main results table. These issues must be resolved before the results can be taken as evidence of a significant advancement.","major_comments":[{"comment":"The paper does not state whether SMOTE is applied inside the cross-validation loop or to the full dataset before the 80/20 split. Figure 1 places \"Data Balancing (SMOTE)\" in the Pre-processing stage, before the Machine Learning stage, and Section 3.3 describes the split and 5-fold CV only afterward. If SMOTE is fit on the entire dataset before splitting, synthetic minority samples generated from test-set originals can leak into the training set, and the classifier is evaluated on data whose synthetic neighbors were seen during training. This would inflate the reported F1 scores for both age and gender estimation, making the central claim unsubstantiated. The authors must clarify the exact pipeline order and, if the current implementation matches the figure, re-run the experiments with SMOTE applied independently to each training fold.","section":"Figure 1; Section 3.3"},{"comment":"The paper reports a single F1 and AUC value per model with no confidence intervals, fold-wise standard deviations, or statistical comparison. Section 3.3 states that 5-fold cross-validation was applied and that performance is the mean across folds, yet no variance information is provided anywhere in the results. Without this, the reader cannot judge whether differences such as XGB 73.26 vs. the ensemble 73.87 in Table 6 are meaningful, nor whether the reported numbers are stable. Provide fold-wise results or at least standard deviations/confidence intervals for the key experiments.","section":"Tables 3-7; Section 3.3"},{"comment":"The abstract states that the Random Forest model achieved the highest F1 score of 77.53% for gender estimation, but Table 6 reports Random Forest at 76.54% and the \"Ensemble of above models\" at 77.53%. Similarly, the abstract credits XGB with the best age estimation F1 of 73.26%, while Table 6 lists the ensemble at 73.87% as the best. This is a direct factual inconsistency in the headline result. The abstract and the results section must be reconciled to report the same model-to-score mapping.","section":"Abstract vs. Table 6"},{"comment":"The baseline results in Table 2 were obtained using all dental features, including canine, first pre-molar, and third molar, which Section 3.2 explicitly excludes from the rest of the study because they were frequently missing. The ensemble results in Tables 6-7 use only the second pre-molar, first molar, and second molar. Consequently, the improvement from baseline to ensemble is confounded by a change in the feature set, not solely by the ensemble method. Either re-run the baseline on the reduced feature set or explicitly discuss this confound when claiming ensemble learning is responsible for the improvement.","section":"Table 2 vs. Section 3.2"},{"comment":"All features are derived from manual landmark placements on radiographs using ImageJ, but the paper provides no inter-observer or intra-observer reliability data and no quantification of landmark placement error. Classifier performance is directly limited by the consistency of these manual measurements; without reliability statistics (e.g., ICC or Bland-Altman limits) on a random subset of radiographs, the reported F1 scores may not generalize to other operators or settings. This is a load-bearing limitation for a method intended as an automated clinical or forensic tool.","section":"Section 3.1.1"}],"minor_comments":[{"comment":"Figure 1 lists SVM and Logistic Regression among the machine learning models, but these models are not discussed in Section 3.3 or in the results tables; conversely, AdaBoost, LGB, and ETC are used later but not listed in Figure 1. Please align the model lists between the figure, the methods, and the results.","section":"Figure 1; Section 4.1"},{"comment":"The abstract describes periapical radiographs, while Section 3.1 repeatedly states that bitewing radiographs were used. This terminology inconsistency should be corrected.","section":"Section 3.1; Abstract"},{"comment":"Table 1 is titled \"Features Extracted\" but the last two rows list Age and Gender, which are the prediction targets rather than features. This should be clarified to avoid confusion.","section":"Table 1"},{"comment":"The relationship between the 80/20 split and the 5-fold cross-validation is unclear: is the 5-fold CV applied to the 80% training set, or to the entire dataset? Please specify the exact evaluation protocol.","section":"Section 3.3"},{"comment":"There are several typographical and wording issues, such as \"evaluation matrix\" in Figure 1 (should be \"evaluation metrics\"), \"Kvp\" for kVp, and \"Amp\" for mA in Section 3.1. A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a practically relevant problem and the authors report institutional ethics approval and a clear measurement protocol. However, the central empirical claims rest on the evaluation pipeline, and the apparent placement of SMOTE before the split is a serious validity threat that cannot be assessed further without code or data. I would encourage the editor to require the authors to clarify and, if necessary, re-run the experiments with leakage-free preprocessing, and to address the abstract inconsistency. The study is not without merit, but the current manuscript is not yet ready for publication without these fixes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an off-the-shelf ML comparison applied to a private dental dataset, and the only genuinely new output is a set of F1 scores for age-group and gender classification using tooth-specific ensembles. Those scores (roughly 73% age, 77% gender) are plausible but not impressive enough to support the 'significant advancement' language in the abstract.\n\nWhat the paper does well: the measurement protocol for CH/CPCH/TCI is described clearly, the related work is thorough, and the authors are upfront about the limitations of previous methods. The tooth-specific ensemble idea—training separate models on second premolar, first molar, and second molar features and combining by vote—is a reasonable variation on standard ensembles, and the non-trainable combiners are a nice sanity check.\n\nThe soft spots are real and need addressing. First, the SMOTE ambiguity: Figure 1 places data balancing before the train/test split, and Section 3.3 does not state explicitly that SMOTE is done only on training folds. If SMOTE was applied to the full dataset before splitting, the test set is partially contaminated with synthetic neighbors, and the reported F1s are optimistic. This needs to be clarified with a precise pipeline statement. Second, no confidence intervals or per-fold variance are reported for any of the tables, so we cannot judge whether the differences between models are meaningful. Third, the baseline in Table 2 uses the full tooth set (canine, first premolar, third molar) that was later excluded; that makes the baseline not directly comparable to the later results. Fourth, there is no inter-observer reliability study for the manual landmarks, which matters because measurement noise would directly degrade classifier performance. Finally, the abstract attributes the 77.53% gender F1 to Random Forest, while Table 6 gives RF 76.54% and reserves 77.53% for the ensemble of all models; the text should be consistent.\n\nNone of these is necessarily fatal. If the authors confirm SMOTE was confined to training data and provide per-fold results, the central numbers may well hold. As it stands, the paper is a modest empirical contribution that would benefit from a serious revision rather than a desk rejection.","headline":"Routine ML on dental x-rays; the reported F1s hinge on an ambiguous SMOTE/split order that a referee should pin down.","tokens_in":13804,"tokens_out":2482,"would_cite":false,"duration_ms":23907,"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":"An ensemble of tooth-specific machine-learning models estimates age group with 73.26% F1 and gender with 77.53% F1 from routine bitewing X-rays.","keywords":["dental biometrics","age estimation","gender estimation","machine learning","ensemble learning","tooth coronal index","explainable AI","bitewing radiographs"],"falsifier":"Take a fresh set of bitewing radiographs and have two or more clinicians independently mark all landmarks, then train the ensemble on one clinician's measurements and evaluate it on the other clinician's: if the age-group or gender F1 drops by more than a few points, or if the landmark measurements themselves have poor inter-observer agreement, the paper's central claim is falsified.","tokens_in":12785,"feed_emoji":"🦷","tokens_out":9404,"duration_ms":84243,"temperature":0.7,"pith_summary":"This paper aims to show that standard dental measurements from routine bitewing X-rays—coronal height, coronal pulp cavity height, and the tooth coronal index derived from them—carry enough information for machine learning to sort living patients into age groups and sexes. The authors test seven classifiers on radiographs of 862 patients and find that a tooth-specific ensemble, where each model learns from a different tooth's measurements and the predictions are combined by voting, reaches F1 scores of 73.26% for five age brackets and 77.53% for gender. The motivation is forensic and clinical: age and gender estimation currently relies on expert visual assessment or invasive tooth extraction, and an automated, explainable system could make the process cheaper, faster, and more consistent. The paper also provides SHAP-based explanations of the driving features, so a dentist can see which measurements mattered for a given prediction.","feed_headline":"Dental X-rays reveal age and sex via voting AI","feed_subtitle":"Per-tooth models voting together hit 73% F1 on age brackets and 78% on sex, no extraction needed.","key_machinery":"The central machinery is the Tooth Coronal Index (TCI), the ratio of coronal pulp cavity height to coronal height multiplied by 100, measured on bitewing radiographs at six tooth positions. The paper trains a separate classifier on each tooth's CH, CPCH, and TCI values and then fuses the classifiers' probabilistic outputs with majority voting or non-trainable combiners (mean, median, maximum, minimum). SHAP (SHapley Additive exPlanations) values rank the contribution of each dental feature to each prediction, providing the explainability layer.","core_discovery":"On its own terms, this paper claims that the three measurements CH, CPCH, and TCI, taken from the second premolar, first molar, and second molar in both jaws of living patients, are sufficient inputs for an ensemble of machine-learning classifiers to estimate age group and gender with F1 scores of 73.26% and 77.53% respectively. The best-performing single models are XGBoost for age and random forest for gender, but the full ensemble of all models, using majority voting or mean/median combiners, matches or exceeds them. The authors interpret this as a step toward replacing subjective expert assessment with a non-invasive, explainable screening tool.","pith_inferences":["The reported accuracy may be optimistic because the landmarks were placed manually with no inter- or intra-observer variability study; a deployment-grade system would need to show the pipeline tolerates measurement noise.","Gender predictions (77.53% F1) beat five-way age predictions (73.26%), hinting that crown-size dimorphism is a stronger signal in this dataset than pulp shrinkage; a testable consequence is that adding root or third-molar measures could close the age gap more than any further classifier tuning.","Collapsing age into five coarse brackets discards the continuous ageing signal; treating age as a regression problem on the same TCI features might yield finer chronological estimates and a more direct comparison with the original TCI regression formulations.","The method's robustness to radiograph quality—older films, varied exposure, different machines—is untested; a realistic benchmark would evaluate the ensemble on a multi-site archive."],"forward_implications":["A dental X-ray already taken for routine care could double as a demography-screening input, adding no extra radiation or tooth extraction.","The same feature set and ensemble design could be tried on skeletal remains, the extension the paper explicitly names as future work.","Because the ensemble outperforms every single model, future development in forensic odontology should expect gains from tooth-specific specialization rather than one global model.","SHAP explanations give clinicians a per-case trace of which dental measurements drove the age or gender call, supporting use where decisions must be justified.","The approach needs only standard periapical radiographs, not CT or laboratory equipment, so it could transfer to low-resource forensic settings."],"supporting_citations":[{"why":"Defines the coronal height and pulp cavity height measurements that are the raw features of the study.","marker":"[10]"},{"why":"Introduces the Tooth Coronal Index used here as the key derived feature for age estimation.","marker":"[9]"},{"why":"Supports the premise that dental radiographs can be used for age and gender determination in living individuals.","marker":"[18]"},{"why":"The image-analysis software used for all landmark measurements and calculations.","marker":"[21]"},{"why":"Provides the SHAP method used to generate the explainability results.","marker":"[16]"},{"why":"Supplies the classifier-combination framework behind the majority-voting and non-trainable combiner ensembles.","marker":"[15]"},{"why":"The XGBoost algorithm that produced the best age-group F1 score.","marker":"[6]"},{"why":"The random forest algorithm that produced the best gender F1 score.","marker":"[5]"}],"fun_headline_variants":["Dental X-rays plus voting AI nail age and sex with SHAP","Explainable ensemble AI for age and sex from tooth X-rays","Voting ML models read teeth: 73% age, 78% sex","Ensemble AI on tooth measurements predicts age and sex with SHAP","Dental AI voting ensemble yields age and sex from X-rays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the manually marked measurements of crown height, pulp cavity height, and the derived ratio are accurate and consistent across observers; the paper presents no inter- or intra-observer reliability data to rule out measurement noise as the true driver of the reported F1 scores.","fun_headline_variants_meta":{"raw":{"variants":["Dental X-rays plus voting AI nail age and sex with SHAP","Explainable ensemble AI for age and sex from tooth X-rays","Voting ML models read teeth: 73% age, 78% sex","Ensemble AI on tooth measurements predicts age and sex with SHAP","Dental AI voting ensemble yields age and sex from X-rays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000647,"raw_usage":{"total_tokens":2998,"prompt_tokens":998,"completion_tokens":2000,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":1907}},"tokens_in":614,"tokens_out":2000,"duration_ms":13338,"temperature":1.0,"reasoning_tokens":1907,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:51:35.472482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fresh set of bitewing radiographs and have two or more clinicians independently mark all landmarks, then train the ensemble on one clinician's measurements and evaluate it on the other clinician's: if the age-group or gender F1 drops by more than a few points, or if the landmark measurements themselves have poor inter-observer agreement, the paper's central claim is falsified.","supporting_citations":[{"cited_title":"Analysis of a random forests model","cited_arxiv_id":null,"evidence_quote":"The random forest algorithm that produced the best gender F1 score."},{"cited_title":"The coronal pulp cavity index: a biomarker for age determination in human adults","cited_arxiv_id":null,"evidence_quote":"Defines the coronal height and pulp cavity height measurements that are the raw features of the study."},{"cited_title":"The coronal pulp cavity index: A forensic tool for age determination in human adults","cited_arxiv_id":null,"evidence_quote":"Introduces the Tooth Coronal Index used here as the key derived feature for age estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the premise that dental radiographs can be used for age and gender determination in living individuals."},{"cited_title":"A unified approach to interpreting model predictions","cited_arxiv_id":null,"evidence_quote":"Provides the SHAP method used to generate the explainability results."},{"cited_title":"On combining classifiers","cited_arxiv_id":null,"evidence_quote":"Supplies the classifier-combination framework behind the majority-voting and non-trainable combiner ensembles."}],"review_version":1}