acceptodds
Under review as a conference paper at ICLR 2027

Leveraging Generative AI for Robust Clinical Predictions: A Case Study on Dental Implant Failures

Abstract

Machine learning has the potential to improve clinical prediction and decision- making, yet its real-world effectiveness is often constrained by the characteris- tics of healthcare data. Clinical datasets are frequently small, heterogeneous, and highly imbalanced, limiting model generalization and reducing the detection of rare but clinically significant outcomes. These challenges motivate the develop- ment of learning frameworks that can effectively leverage limited data while pre- serving the statistical characteristics necessary for robust prediction. This paper investigates synthetic tabular data generation as a strategy for improving predic- tive modeling under data scarcity and class imbalance. We propose GC-TabGAN (Gaussian-Conditional Tabular GAN), a customized Generative Adversarial Net- work (GAN) for generating realistic synthetic clinical records, and integrate it with a Feed Forward Neural Network (FFNNet Baseline) for downstream pre- diction. Dental implant failure is used as a representative clinical case study be- cause implant failures are relatively rare, arise from complex interactions among patient, clinical, and procedural factors, and therefore present a challenging pre- diction task. Using a public dataset of 747 implants, of which 117 (15.7%) failed, and a 10-fold cross-validation protocol in which the generator is refit within every training fold, we add 5%, 25%, 50%, and 90% of a 1,000-record fold-specific syn- thetic pool to the real training data. Synthetic augmentation mainly improves the detection of implant failures. Using only real data, the FFNNet Baseline achieved an accuracy of 0.92, an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.81, and a sensitivity to implant failure of 0.64. Augmenting the train- ing data with synthetic samples raised the sensitivity to 0.76–0.80 while maintain- ing a specificity of 0.97–0.98; with 50% synthetic samples, the model reached an accuracy of 0.95, an AUC of 0.89. In contrast, oversampling the training data with SMOTE raised sensitivity to 0.88 but reduced specificity to 0.66, with an AUC of 0.77, well below that of GC-TabGAN, indicating that GC-TabGAN improves dis- crimination rather than merely shifting the decision boundary toward the minority class. These findings suggest that distribution-aware synthetic tabular data can partly offset data scarcity and class imbalance in small clinical datasets, a setting common in healthcare where large, balanced datasets are difficult to collect.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
dental implants failure, predictive modeling, machine learning, generative AI, conditional GAN

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AEMrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method for synthetic tabular data generation using GANs, and demonstrates that combining this synthetic data with real data results in an improvement in performance on a particular clinical application (detecting dental implant failures). For instance, it raises the AUC of their baseline method from 0.81 to 0.89 with 50% synthetic examples.

Critical strengths

The prose is genuinely well-written and the paper is easy to follow, in contrast to many of the AI-slop papers in my reviewing batch. Kudos to the authors for writing their paper the good-old-fashioned way, or at least making it appear that way :)

Overall, the paper presents a clearly-explained approach to the problem, and conducts a reasonably thorough evaluation of their approach (with plenty of well-motivated baselines) on this particular dataset. I appreciated the care that went into designing the synthetic data generation approach to avoid potential leakage (via cross-validation, which seems computationally expensive, but at least principled), and Figure 2 presents a rare sight: An ROC curve that shows different trade-offs at different levels (e.g., for a very low FPR, the baseline might be a better choice, whereas their method works better in the high-recall regime).

Critical weaknesses

The obvious limitations of this paper are the narrow focus and scope, showing improvement in performance on a specific clinical dataset. While the method is fairly general, it is unclear how we should expect the method to generalize to other tabular prediction tasks. The authors may want to consider using TabArena or other compilations of tabular prediction tasks to see if their method works more generally, and to try and derive some insights into why it seems to work.

Moreover, insofar as the goal of the paper is to explore synthetic data generation, there do not appear to be any other baselines that perform synthetic data generation: The baselines are largely either (a) classical classification methods, and (b) the FFNNet baseline + SMOTE, which is "a reference for conventional imbalance handling". I raise this point in part because I was surprised to see the use of GANs as opposed to diffusion models or other synthetic data-generation approaches, given the difficulties and instabilities typically encountered while training GANs. Perhaps the GAN training is simpler in this case, given the fact that all features modeled by the GAN are first tested for being continuous and Gaussian? But in that case, would it work equally well to just directly fit via MLE a multi-variate Gaussian to all of the Gaussian features, as opposed to using a GAN?

Finally, I did have some minor concerns related to the restrictions imposed by the synthetic data generation model: First, the testing of continuous features for Gaussianity seems somewhat arbitrary (few variables will be exactly Gaussian), and I wonder how sensitive the results are to that design choice. Moreover, I read Equation (3) as saying that the approach requires fitting a MLE estimate to the Gaussian distribution for each level of every non-Gaussian feature, which seems like it could be quite computationally intensive if there are a lot of non-Gaussian features. Finally, my understanding of Equation (3) is that the non-Gaussian features as modeled as conditionally independent given the Gaussian features: For instance, if all features were discrete, would this approach reduce to treating all features as independent? That seems like an unrealistic restriction.

Questions for authors

Can you answer my questions below to clarify how the model works?

  • Does Equation (3) fit the conditional distribution separately for each level of each non-Gaussian feature? E.g., if I have three binary non-Gaussian features x_1, x_2, x_3, do I have six conditional fits (one for x_1 = 0, one for x_1 = 1, etc)?
  • Am I right to read Equation (4) as modelling each non-Gaussian feature as conditionally independent given the Gaussian features? If so, am I correct to infer that if there are no Gaussian features, then each non-Gaussian feature is sampled independently?
Reviewer_vgnrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether synthetic tabular data can improve dental implant failure prediction under limited and imbalanced clinical data. It proposes GC-TabGAN, which uses a GAN to generate approximately Gaussian features and a Gaussian conditional model to sample categorical features. Separate generators are trained for successful and failed implants, and the generated records are used to augment the training data of a feed-forward neural network. The results show improvements in failure detection compared with training on real data alone. The method also performs better than standard SMOTE.

Critical strengths
  1. The paper addresses a relevant problem in clinical prediction. The paper focuses on failure sensitivity rather than only reporting overall accuracy. The proposed method also handles continuous and categorical variables differently, which is reasonable for clinical tabular data.
  2. The authors fit the synthetic data generator within each cross-validation training fold and evaluate the classifier only on real held-out records. The reported improvements in failure sensitivity are encouraging because specificity remains relatively high across the augmentation settings.
Critical weaknesses
  1. The AUC calculation needs clarification. The AUC values reported for the FFNNet baseline, SMOTE, and GC-TabGAN configurations in Table 2 match the average of sensitivity and specificity after rounding. Whether AUC is calculated using binary predictions rather than continuous prediction scores. If binary predictions are used, the reported AUC would effectively correspond to balanced accuracy at a single threshold. The paper uses the AUC improvements to argue that GC-TabGAN improves discrimination rather than simply changing the classification boundary.
  2. The paper mainly compares GC-TabGAN with standard SMOTE and the real-data baseline. However, the augmentation methods use different class proportions, and the comparison is conducted at a fixed classification threshold. Since many features are categorical, standard SMOTE may not be the most appropriate baseline. Comparisons with random oversampling, class weighting, SMOTENC, or other tabular generators would be more informative. The evaluation is also limited to one dataset and a single random seed. With only 117 failure cases, the reported improvements may be sensitive to the data split. Repeated evaluation and paired statistical analysis would help determine how reliable these gains are.
Questions for authors
  1. How exactly is AUC calculated in Table 2? The reported values appear to match balanced accuracy for every proposed-model configuration. Are continuous prediction probabilities or binary class labels used? Please report ROC-AUC computed from held-out prediction scores and clarify how the ROC curves in Figure 2 are generated.
  2. How to provide comparisons with class weighting, random oversampling, or SMOTENC using comparable class proportions? It would also be helpful to report results across multiple random seeds or repeated cross-validation, together with uncertainty estimates for the improvement over the real-data baseline.
Reviewer_hYfKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes GC-TabGAN, a synthetic tabular data generation method designed to improve clinical predictions in oral healthcare, specifically dental implant failure prediction. The approach separates Gaussian and non-Gaussian features during data generation and uses the generated samples to augment the training data of a neural network classifier. The authors evaluate the method using 10-fold cross-validation, generating synthetic data independently within each training fold to prevent data leakage. The results suggest improvements in failure detection compared to training on real data alone and to SMOTE-based augmentation.

Critical strengths

A positive aspect is the attention given to preventing data leakage, as synthetic data generation is performed independently within each training fold during cross-validation.

Critical weaknesses

Line 054: The transition between the discussion of clinical risk factors and the subsequent literature review is not sufficiently clear. The sentence appears to introduce a new argument, but its connection to the preceding discussion could be improved.

The research questions are relatively limited in scope. They focus primarily on whether augmentation improves predictive performance and how performance changes with the amount of synthetic data. However, they do not directly investigate the quality of the generated data, the preservation of clinical feature dependencies, or the advantages of the proposed architecture over existing generative approaches.

The work lacks comparisons with other synthetic data generation methods. Although several techniques are discussed in the related work, the experimental evaluation only includes SMOTE as an augmentation baseline. This makes it difficult to assess whether the proposed architecture offers meaningful advantages over established tabular generators.

The authors should better explain what constitutes an approximately Gaussian feature in this context and clarify the implications of using a normality test to determine how each feature is modeled.

The motivation for separating Gaussian and non-Gaussian features should be explained more thoroughly. In particular, the authors should clarify why this design is preferable to modeling the joint distribution directly.

The proposed approach should be compared with a conventional GAN or an established tabular GAN under the same experimental protocol. This would help determine whether the proposed Gaussian-conditional component provides an actual advantage.

Line 203: The mention of standardization, normality testing, and discretization suggests that the previous dataset description is incomplete. These preprocessing steps are important to understanding the methodology and should be clearly described as part of the data preparation procedure.

The authors propose a synthetic data generation technique, but the main downstream evaluation is also performed using a neural network architecture developed by the authors. Although Table 2 includes results from other classifiers, those results are taken from previous studies and do not evaluate the proposed augmentation strategy. The authors should demonstrate whether GC-TabGAN also improves the performance of other classification models.

Lines 234–239: The authors should better justify the architectural choices of the generator and discriminator, including the use of multilayer perceptrons, instance noise, and the selected training configuration. It is unclear whether these choices were empirically validated or primarily adopted for implementation simplicity.

Figure 2: The ROC curves do not show uniform improvements across all operating points. Although the augmented models achieve higher overall AUC, the curves intersect in some regions. Therefore, claims about improved performance should distinguish between overall discrimination and improvements at specific decision thresholds.

The authors interpret the results as evidence that GC-TabGAN improves discrimination rather than merely addressing class imbalance. However, SMOTE and GC-TabGAN are evaluated under different class-balancing conditions. While SMOTE fully balances the classes, GC-TabGAN produces a substantially smaller change in the class distribution. This introduces a potential confounding factor, making it difficult to isolate whether the observed performance differences are due to the data generation method or the degree of class rebalancing. The authors should compare both methods under equivalent class distributions and consider threshold tuning to ensure a fair comparison. Although the AUC results suggest that GC-TabGAN may provide additional benefits, the current experimental design does not fully isolate the contribution of the proposed generation strategy.

The paper does not fully address the research questions introduced in the introduction. While the experiments examine predictive performance across different augmentation ratios, the evaluation is insufficient to establish whether the proposed generative approach provides advantages over alternative generators or consistently preserves the characteristics of the original clinical data.

There are important limitations in the experimental design. In particular, the absence of comparisons with alternative generators, the limited evaluation of synthetic data quality, and the reliance on a single dataset and random seed make it difficult to assess the robustness and generalizability of the proposed approach.

Overall, the work presents an interesting application of synthetic data augmentation to a clinically relevant prediction problem. However, the methodology and experimental evaluation appear incomplete, and the organization could be improved. Several important design decisions are insufficiently justified, while the current experimental evidence does not adequately isolate the contribution of the proposed generative architecture.

Questions for authors

Does separately modeling Gaussian and non-Gaussian features affect the quality and consistency of the generated records? Clinical variables can exhibit strong dependencies, and separating their generation may fail to preserve some of these relationships. Although the proposed method conditions categorical variables on generated Gaussian features, it samples the categorical variables independently given those features. The authors should investigate whether this assumption preserves clinically relevant dependencies.

How do the authors ensure that the generated features form coherent synthetic patient records? The method combines generated Gaussian features with conditionally sampled categorical features, but it is unclear whether the resulting records preserve the joint relationships observed in the original data. An evaluation of feature correlations, conditional dependencies, and clinically plausible combinations would strengthen the methodology.

Why was SMOTE selected as the only augmentation baseline? Given that the paper discusses several generative architectures, comparisons with established synthetic tabular data generators would provide a more meaningful evaluation of the proposed method.

Table 2: Standard deviations are reported only for the authors' models, while the remaining results are taken from previous studies. Were all models evaluated under comparable training and evaluation protocols? The authors acknowledge differences in cross-validation strategies, which makes direct comparison difficult. A controlled evaluation using the same splits, preprocessing, and metrics would provide stronger evidence.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.