{"id":"5bbfb6cd-e41f-4141-a218-1ed42184c9c5","arxiv_id":"2506.21927","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN-LSTM model fitted to 40 quarterly sales records of a single Egyptian cancer drug reports lower MSE and RMSE than CNN, LSTM, and RNN on an unspecified test split.","lead":"This paper applies a standard CNN-LSTM neural network to forecast quarterly sales of one cancer drug in Egypt from 2015 to 2024, reporting MSE 1.150 and RMSE 1.072. A generalist should care only slightly: the dataset is 40 points, the method is routine, and no code or data is provided, so the result cannot be checked.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central forecasting claim is unsupported because the paper never specifies the train/test split for the 40 quarterly observations, so the reported test-set MSE/RMSE (Section 5, Table 1) cannot be distinguished from in-sample fit.","rationale":"The reader identified the missing train/test split as the weakest assumption; I agree this is the single most load-bearing issue. I considered other possible concerns, such as the mismatch between the model's capacity and the 40-point dataset, and the lack of error bars or statistical significance; however, these all become secondary once the evaluation protocol is unspecified, because no metric can be interpreted. A concrete check would be to demand the split and code; this settles the concern because either the split exists and can be reproduced, or it does not and the paper's claim is unverifiable. I also note that even with a split, the small sample size would warrant a naive baseline. The reader's REJECT verdict remains appropriate.","tokens_in":894,"tokens_out":636,"duration_ms":35767,"concrete_test":"Ask the authors to provide the exact split (e.g., 'first 30 quarters training, last 10 testing'), random seed, and the code or predictions per quarter. Then recompute MSE/RMSE for CNN-LSTM and the three baselines on the same split, and add a naive persistence baseline (predicting the previous quarter's sales) and a seasonal naive baseline. If the CNN-LSTM does not beat persistence, or if the authors cannot supply the split details, the central claim of superior forecasting performance fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's entire conclusion (Section 5) rests on the reported test-set MSE of 1.150 and RMSE of 1.072. To interpret these as out-of-sample forecasts, one must know how the 40 quarterly observations (Section 3) were partitioned into training and test sets, including whether any points were held out, the split ratio, and the random seed. The paper provides none of this; it only states 'test set' without defining it. Given the model's capacity (two CNN layers with 64 and 128 filters, two LSTM layers with 128 units, dropout 0.3), the model has tens of thousands of parameters yet only roughly 30 training samples if a typical split is used, making severe overfitting likely. Without the split specification, the reported metrics are equally consistent with in-sample evaluation, a leaky preprocessing pipeline (e.g., normalization computed on the full series before splitting), or a single favorable run. Consequently, the comparative advantage over CNN, LSTM, and RNN in Table 1 is not established. The burden is on the authors to show the evaluation is genuinely out-of-sample; the manuscript as written does not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a CNN-LSTM hybrid model for forecasting quarterly sales volume of a single cancer drug in Egypt over 2015-2024 (40 quarters). The authors report an MSE of 1.150 and RMSE of 1.072, which they attribute to a test set, and compare these numbers with those of CNN, LSTM, and RNN baselines. The central claim is that the hybrid architecture outperforms the single models and provides a reliable data-driven tool for pharmaceutical forecasting.","tokens_in":6541,"tokens_out":3153,"duration_ms":37806,"significance":"If the reported metrics were genuinely out-of-sample and reproducible, the study would provide a useful empirical data point on hybrid deep learning for pharmaceutical demand forecasting on small, real-world data. The paper's stated contributions---a multidimensional quarterly dataset, a CNN-LSTM architecture, and a benchmark against single models---are plausible but currently unverifiable. The manuscript provides no dataset, no code, no error bars, and no explicit train/test protocol. The significance is therefore prospective rather than demonstrated: the architecture choice is reasonable, but the evidence does not currently support the forecasting claim.","major_comments":[{"comment":"The central claim that the CNN-LSTM model 'performs well on the test set' is unsupported because the paper never defines the train/test split. Section 3 describes the data as 40 quarterly observations and mentions preprocessing, but no partition ratio, chronological cutoff, random seed, or cross-validation procedure is reported. With only 40 data points, the quoted MSE of 1.150 and RMSE of 1.072 are equally consistent with in-sample fitting, a single favorable run, or a leaky preprocessing pipeline. This is load-bearing because the Abstract, Section 5, and the Conclusions all rest on these two numbers. The authors must specify how the test set was constructed and demonstrate that the reported errors are out-of-sample.","section":"§5, Table 1 (Comparison of experimental results)"},{"comment":"The model capacity is very large relative to the sample size: two 1D-CNN layers with 64 and 128 filters, two LSTM layers with 128 hidden units, and stateful training. With roughly 40 quarterly observations, the effective training sample after any plausible split is maybe 25-30 points. The paper reports only a dropout rate of 0.3 and gives no information about epochs, early stopping, optimizer, learning rate, batch size, or how many runs were averaged. Without this information and without a proper train/test separation, the statement in the Conclusions that the model shows 'strong generalization and stability' is not supported.","section":"§4 (CNN-LSTM Model Introduction)"},{"comment":"The comparison with CNN, LSTM, and RNN baselines is uninterpretable. No architectural details, training configurations, or evaluation protocols are given for the baselines beyond generic descriptions. The table lists a single MSE and RMSE per model with no error bars, no repeated runs, and no significance test. On a 40-point series, run-to-run variability can easily change these numbers by a large margin; the text's claim of 'significant improvement' is therefore not justified by the evidence presented.","section":"Table 1 (Comparison of experimental results)"},{"comment":"Reproducibility is not possible from the manuscript: the dataset is not made available, the scraping sources are not listed in enough detail to re-acquire the data, and no code is provided. In addition, the preprocessing description says 'unit normalization and time alignment were performed on numeric fields like price and sales volume' before any mention of splitting; if normalizing statistics are computed on the full series, test information leaks into the training process. The authors must state that normalization parameters are fitted only on the training partition, and if that is already the case, they should say so explicitly.","section":"§3 (Data Introduction)"}],"minor_comments":[{"comment":"The paper contains two different tables both labelled 'Table 1' (the variable description table and the model comparison table); the second should be renumbered and the in-text references corrected.","section":"Throughout"},{"comment":"The percentages in the regional distribution do not sum to 100% (four regions at 20% each plus Europe at 19.9% totals 99.9%); please correct the numbers or the description.","section":"Figure 1"},{"comment":"There are two section 5 headings: 'Model result analysis' and 'Conclusions'; the Conclusions should be renumbered as a separate section.","section":"§5 (heading numbering)"},{"comment":"The forecast curve would be far more informative with labeled axes, units, a legend distinguishing actual and predicted values, and ideally a shaded confidence interval or at least the residual series; the present figure supports only a qualitative reading.","section":"Figure 3"},{"comment":"The variable 'Date' is described as 'the timestamp when the data was scraped', yet the data are quarterly sales records spanning 2015-2024; please clarify whether the timestamp refers to the calendar quarter of sales or to the time of collection, as these are materially different for time-series modeling.","section":"Table 1 variables"},{"comment":"Reference [15] contains a corrupted author entry ('** Y'), and several references lack complete page ranges or DOIs; please correct these in a final copyediting pass.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper has a clearly stated central claim, but the evaluation protocol is missing at the most load-bearing point: the test set is never defined for a 40-observation series. The reported accuracy metrics could be in-sample fits, so the contribution cannot be validated by any reader. The issues are not merely editorial; they require a new experimental design with explicit splits, repeated runs, and ideally access to data and code. In the current scope of the manuscript, that is not a minor fix. However, I would not rule out a substantially rewritten version with a proper out-of-sample evaluation and availability of the dataset."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a routine CNN-LSTM application to 40 quarterly sales records of one cancer drug in Egypt. The reported test-set MSE of 1.150 and RMSE of 1.072 are the entire basis for the forecasting claim, but the paper never says how the 40 quarters were split into train and test. Those numbers could just as easily be in-sample fits. That alone sinks the central claim.\n\nTo its credit, the paper does a few things right. The literature review touches relevant work on pharmaceutical sales forecasting with LSTM and XGBoost, and the model description is reasonably concrete: two 1D CNN layers with kernel sizes 3 and 5 and 64/128 filters, followed by two LSTM layers with 128 units and dropout 0.3. It also compares against CNN, LSTM, and RNN baselines, and it admits limitations—no external variables, single region, need for multi-region validation. So the authors are not oblivious to the bigger picture.\n\nBut the soft spots are load-bearing. First, the missing train/test split. With 40 quarterly observations and a model that likely has tens of thousands of parameters, overfitting is almost guaranteed. Second, there are no error bars or repeated runs; the baselines show LSTM at MSE 1.956, which is far better than CNN but still worse than the hybrid—yet without variance or significance testing, that ordering could be noise. Third, the data are proprietary and not made available, so nothing is independently checkable. Fourth, the paper does not say which variables are used as inputs versus the target. It lists drug type, company, price, effectiveness, and user reviews; we don't know how these are encoded or whether the model saw them at all. Finally, the dataset itself is questionable—web-scraped from 'several reputable Egyptian pharmaceutical websites' with variables like 'Effectiveness' and 'UserEvaluate' that have no clear measurement scale.\n\nIn short, the paper is a straightforward application of a standard architecture to a tiny dataset, and the evaluation is too under-specified to support the headline result. It could be a useful teaching example of how not to report time-series experiments, but it is not a contribution to forecasting methodology or to cancer drug sales analysis.\n\nRecommendation: desk reject. If the authors later provide code, data, and a clear train/test protocol, a re-submission could be worth a second look, but not as it stands.\n\nRegards,","headline":"A routine CNN-LSTM application to 40 quarters of one drug's sales; the reported accuracy is unverifiable because the train/test split is never defined.","tokens_in":7019,"tokens_out":2647,"would_cite":false,"duration_ms":26254,"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 CNN-LSTM hybrid forecasts quarterly cancer drug sales with lower error than CNN, LSTM, or RNN on the same Egyptian dataset.","keywords":["CNN-LSTM","time series forecasting","cancer drug sales","sales volume prediction","pharmaceutical supply chain","deep learning","MSE","RMSE"],"falsifier":"Re-run the comparison with a stated split of the 40 quarters (for example, the first 32 as training and the last 8 as held-out test) across multiple random seeds; if CNN-LSTM does not beat LSTM on the held-out test MSE, the central forecasting claim fails.","tokens_in":6123,"feed_emoji":"💊","tokens_out":6251,"duration_ms":56668,"temperature":0.7,"pith_summary":"This paper tries to establish that a hybrid deep learning architecture, CNN-LSTM, forecasts quarterly sales of a specific cancer drug in Egypt more accurately than CNN, LSTM, or RNN models applied to the same data. The authors assemble 40 quarterly records from 2015 to 2024, with features such as price, drug type, company, effectiveness, and sales volume. On their test results, CNN-LSTM reaches an MSE of 1.150 and RMSE of 1.072, compared with an MSE of 3.526 for CNN and 1.956 for LSTM. The practical motivation is that better sales forecasts help pharmaceutical companies plan production, manage inventory, and support healthcare policy decisions.","feed_headline":"CNN-LSTM cuts forecast error on cancer drug sales","feed_subtitle":"On 40 quarters of Egypt sales data, the hybrid model beats CNN, LSTM, and RNN on both MSE and RMSE.","key_machinery":"The central object is the CNN-LSTM hybrid network: two one-dimensional convolutional layers (kernel sizes 3 and 5, with 64 and 128 filters, batch normalization, ReLU, and max-pooling) extract local temporal features from the multi-channel sales series, and two stacked LSTM layers (128 hidden units each, dropout 0.3, stateful with reset) capture long-term dependencies. A fully connected dense layer produces the next quarter's sales volume regression output. The machinery's role is to fuse short-window patterns with long-range trends so the model can track both seasonal fluctuations and gradual growth in a volatile series.","core_discovery":"The central claim is that combining a convolutional front end with a long short-term memory back end captures both local temporal patterns and long-term dependencies in cancer drug sales data, and that this combination outperforms each component alone. In the reported comparison, CNN-LSTM achieves MSE 1.150 and RMSE 1.072 on the test set, beating CNN (3.526, 1.878), LSTM (1.956, 1.399), and RNN (2.026, 1.423). The author's interpretation is that CNN extracts local features such as price fluctuations and sales peaks while LSTM models seasonal and inter-quarter trends, and that the fused representation yields the best forecast. The paper claims this supports data-driven decision-making for pharmaceutical production, inventory, and healthcare policy.","pith_inferences":["Because the paper does not report the train/test split, the quoted test errors should be treated as conditional on an unspecified split, and a proper validation design is required before the model ranking is accepted.","The single-drug, single-region, 40-quarter scope leaves the architecture's generality untested; applying the same pipeline to multi-drug or multi-region data would be a direct way to check it.","The paper's own limitation list suggests a concrete extension: adding external shock variables such as policy changes, competing products, and public health events could reduce the deviations seen during sudden market shifts.","An ablation test that replaces the CNN block with dense layers of comparable size before the LSTM would isolate how much of the reported gain actually comes from convolutional feature extraction."],"forward_implications":["On the same 40-quarter Egypt dataset, CNN-LSTM lowers MSE by about 2.38 relative to CNN and by about 0.81 relative to LSTM.","The trained pipeline can produce next-quarter sales volume forecasts for this drug, which can feed directly into production planning and inventory management.","The reported alignment of predicted and actual sales curves implies the model captures seasonal fluctuations and long-term growth trends in this series.","The multidimensional feature set (price, effectiveness, company, drug form, user evaluation) can be reused as inputs for the same architecture on similar pharmaceutical sales data."],"supporting_citations":[{"why":"Provides the convolutional neural network component used for local temporal feature extraction in the hybrid architecture.","marker":"[3]"},{"why":"Provides the LSTM component used to model long-term dependencies in the sales sequence.","marker":"[4]"},{"why":"Prior work combining conventional and machine-learning models for pharmaceutical sales prediction, the baseline the hybrid approach extends.","marker":"[6]"},{"why":"Prior deep-learning drug sales forecasting that found LSTM outperforms MLP and CNN, motivating the hybrid comparison.","marker":"[7]"},{"why":"A 1D-CNN-LSTM hybrid neural network application that supplies the architectural pattern for combining one-dimensional convolution with LSTM.","marker":"[10]"},{"why":"The original LSTM formulation that underlies the two stacked LSTM layers with gating mechanisms.","marker":"[14]"},{"why":"An interpretable CNN-LSTM forecasting model with feature engineering, precedent for using the hybrid for health-data time series.","marker":"[15]"}],"fun_headline_variants":["CNN-LSTM outperforms rivals on cancer drug sales","Hybrid CNN-LSTM beats single models in drug forecast","Cancer drug sales predicted better by CNN-LSTM","CNN-LSTM improves accuracy on volatile drug sales","Deep learning hybrid wins on cancer sales prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the quoted MSE and RMSE are out-of-sample test errors, yet the paper never specifies the train/test split, random seed, or cross-validation for the 40 quarterly observations.","fun_headline_variants_meta":{"raw":{"variants":["CNN-LSTM outperforms rivals on cancer drug sales","Hybrid CNN-LSTM beats single models in drug forecast","Cancer drug sales predicted better by CNN-LSTM","CNN-LSTM improves accuracy on volatile drug sales","Deep learning hybrid wins on cancer sales prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1236,"prompt_tokens":962,"completion_tokens":274,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":201}},"tokens_in":578,"tokens_out":274,"duration_ms":3297,"temperature":1.0,"reasoning_tokens":201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:14:30.177188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison with a stated split of the 40 quarters (for example, the first 32 as training and the last 8 as held-out test) across multiple random seeds; if CNN-LSTM does not beat LSTM on the held-out test MSE, the central forecasting claim fails.","supporting_citations":[{"cited_title":"Breath analysis system with convolutional neural network (CNN) for early detection of lung cancer[J]","cited_arxiv_id":null,"evidence_quote":"Provides the convolutional neural network component used for local temporal feature extraction in the hybrid architecture."},{"cited_title":"Comparative study of long short -term memory (LSTM), bidirectional LSTM, and traditional machine learning approaches for energy consumption prediction[J]","cited_arxiv_id":null,"evidence_quote":"Provides the LSTM component used to model long-term dependencies in the sales sequence."},{"cited_title":"Applying Machine Learning and Statistical Forecasting Methods for Enhancing Pharmaceutical Sales Predictions[J]","cited_arxiv_id":null,"evidence_quote":"Prior work combining conventional and machine-learning models for pharmaceutical sales prediction, the baseline the hybrid approach extends."},{"cited_title":"Drug sales forecasting in the pharmaceutical market using deep neural network algorithms[J]","cited_arxiv_id":null,"evidence_quote":"Prior deep-learning drug sales forecasting that found LSTM outperforms MLP and CNN, motivating the hybrid comparison."},{"cited_title":"Wear prediction of high performance rolling bearing based on 1D-CNN-LSTM hybrid neural network under deep learning[J]","cited_arxiv_id":null,"evidence_quote":"A 1D-CNN-LSTM hybrid neural network application that supplies the architectural pattern for combining one-dimensional convolution with LSTM."},{"cited_title":"Long short -term memory[J]","cited_arxiv_id":null,"evidence_quote":"The original LSTM formulation that underlies the two stacked LSTM layers with gating mechanisms."},{"cited_title":"COVID -19 vaccine prediction based on an interpretable CNN-LSTM model with three -stage feature engineering[J]","cited_arxiv_id":null,"evidence_quote":"An interpretable CNN-LSTM forecasting model with feature engineering, precedent for using the hybrid for health-data time series."}],"review_version":1}