Disease diagnosis and prediction using deep learning: a review

Overview of machine learning

ML, a branch of AI, allows computers to “self-learn” from training data and improve over time without explicit programming. Detecting patterns in data and learning from them permits machine learning algorithms to develop their predictions. Medical experts can apply machine learning in healthcare to create better diagnostic tools for examining medical images. For instance, medical imaging (such as X-rays or MRI scans) can utilize a machine-learning algorithm for pattern recognition. Machines learn on their own without being programmed by humans. Discovering patterns and learning from ML are useful in healthcare research for better diagnostic tools and for examining images. Machine learning can address multiple disease diagnoses and prevention. ML is used in the healthcare industry in many ways, such as diagnostics, treatment, research, drug discovery, and healthcare administration.

Figure 1 shows the ML models, which are classified into supervised learning, unsupervised learning, reinforcement learning, transfer learning, and federated learning. This organized overview contextualizes the methodological underpinnings on which the study’s proposed model is built, guiding readers through the reasoning behind the selection of various methodologies. In machine learning, supervised learning involves using labeled datasets to train algorithms to recognize patterns and predict outcomes (Jiang, Gradus & Rosellini, 2020). In unsupervised learning, the training datasets are not used to supervise the models throughout the machine-learning process. Instead, the models extract the insights and hidden patterns from the provided data (DataRobot, 2021).

Machine learning techniques for disease diagnosis

New methods of disease diagnosis in medicine are being developed using ML. Figure 1 shows the number of techniques that can be used to diagnose the disease accurately, and depending on the disease, the best algorithm can be used. Here are a few related works. Heart disease is a significant health problem that affects millions of people worldwide. Early and accurate diagnosis of heart disease is essential for adequate treatment and prevention. Parkinson’s disease is a neurodegenerative disease brought on by the death of brain cells that produce dopamine. One interesting strategy being developed by researchers for the early diagnosis of Parkinson’s disease is using machine learning. A machine learning-based method created by Senturk (2020) successfully classified Parkinson’s patients with an accuracy of 93.84%. It showed that machine learning might be a valuable technique for the early diagnosis of Parkinson’s disease. Based on ML theory, notably Support Vector Machines (SVM) and Random Forests (RF), Huang, Gao & Ye (2021) created an intelligent data-driven model. The performance and accuracy of Cough Variant Asthma (CVA) diagnosis can be enhanced. Additionally, it was demonstrated that the suggested methodology was a simple way to increase the effectiveness of disease diagnosis. ML is effective in predicting several diseases, including diabetes, cancer, and heart disease, and ML algorithms are as accurate as or even more accurate than human doctors. Early diagnosis is critical for the successful treatment of many diseases. ML has the potential to diagnose diseases early when they are more curable. This could lead to improved patient outcomes and reduced healthcare costs.

Using spirometry data, Bhattacharjee et al. (2022) created machine-learning models to categorize lung disorders into obstructive and non-obstructive categories. Using 5-fold Cross-Validation (CV), models were trained using spirometry data from 1,163 patients. An additional blind dataset of 151 patients was used for external validation. With an accuracy of 83.7%, the Multi-Layer Perceptrons (MLP) model operated at peak efficiency. In summary, despite the many challenges that need to be solved, recent advances in ML have led to new challenges in the medical field, like unbalanced data, ML interpretation, and ML ethics (Ahsan et al., 2020). Deep Learning is a potential approach for disease detection and prediction, with ANN serving as the fundamental notion of machine learning. DL models can recognize intricate links and patterns in data. Advanced DL architectures like CNN and RNN can efficiently process large volumes of data in real-time.

The comparative analysis in Table 1 indicates that while each research work significantly adds to its field, there is evident fragmentation in methodological integration. It emphasizes the essential need to include interpretability, multimodal data, and empirical validation to enhance the therapeutic relevance of deep learning models. Few research studies have concurrently examined essential AI components, including explainability, privacy, generalizability, and real-world scalability. This suggests that contemporary research often prioritizes performance criteria above configuration with the broader requirements of clinical implementation. Notwithstanding sophisticated models, no comprehensive research integrates large-scale data utilization, explainable AI, Transformer models, class balancing, and federated learning elements essential for ethically responsible and transparent AI in healthcare. This indicates a considerable deficiency in cohesive, comprehensive, intelligent diagnostic systems. Consequently, there is an urgent need for multidisciplinary and multimodal AI pipelines that exhibit high performance while being interpretable, secure, and relevant to various patient demographics.

Convolutional neural networks

Convolutional Neural Networks are deep learning models used in image recognition, video analysis, Natural Language Processing (NLP), medical imaging, remote sensing, generative tasks, and speech recognition, and they need a large set of input parameters. CNNs are used for automatic feature extraction and operate well on large data sizes. CNNs are used in medical image analysis for disease diagnosis and prediction and mathematical notation is outlined in Eq. (1). Compared to MultiLayer Perceptron (MLP)s, CNNs require less memory and time to train on the data. CNN architecture contains layers such as convolutional, pooling, fully connected, and output.

Figure 3 depicts the architecture of CNNs, which have proven essential for disease detection via deep learning. It demonstrates how CNNs systematically extract and abstract spatial elements from medical images, emulating clinical pattern recognition for disease diagnosis. Each layer represents physiologically significant modifications that facilitate precise categorization. The stratified architecture of CNNs facilitates a gradual and autonomous learning process from unprocessed medical images to precise diagnostic predictions. Wang (2024) used a CNN to diagnose breast cancer from X-ray images and achieved an accuracy of 92.1%. The result highlights the capability of deep learning to aid radiologists by offering a second opinion or automating preliminary screening, thereby alleviating the burden and enhancing early detection rates.

In the study of medical images for disease diagnosis and prognosis, CNNs have produced encouraging results. To extract information from images and learn a hierarchical representation of the image, CNNs employ convolutional layers. CNNs have been used to diagnose several disorders, including pulmonary, cardiovascular, and cancer. Sudhish, Nair & Shailesh (2024) incorporated CNNs into the Content-Based Medical Image Retrieval (CBMIR) framework, providing a revolutionary method for disease identification and facilitating the automated, efficient, and precise retrieval of pertinent medical images. The suggested pipeline utilizes hierarchical feature extraction, multi-level gain-based selection, and sophisticated indexing algorithms to achieve optimal efficiency in managing sizeable medical image libraries, effectively tackling significant.

Recurrent neural networks

RNNs are useful for tasks requiring sequential data, such as time series or electronic health records. RNN performs better than other DL models when healthcare data includes time series properties, as shown in Fig. 4. Each unit processes the input sequentially while retaining prior context, making it effective for analyzing time-dependent clinical data, such as ECG, Electroencephalogram, or patient history and mathematical notation is outlined in Eqs. (1), (2).

Hidden state update:

(1) $$h_t=f(W_{hh}{h}_{t-1}+W_{xh}{x}_t+b_h).$$

Output:

(2) $$y_t=W_{hy}{h}_t+b_y.$$

RNNs are widely used to predict changes in healthcare conditions when the variables exhibit a time series aspect. They have been utilized to predict the course of disease and the effectiveness of treatment. CNNs have extracted the geographical distribution of skin lesions to help identify skin cancer.

By examining the temporal trends in heart rate variability, RNNs have been utilized to detect cardiac disease. The spatial distribution of blood glucose levels has been analyzed using CNNs to identify diabetes. RNNs have been used to identify Alzheimer’s disease by examining the temporal patterns of brain activity. CNNs have studied the geographical distribution of skin lesions to help identify skin cancer. By reviewing the temporal trends in heart rate variability, RNNs have been utilized to detect cardiac disease. CNNs have examined the spatial distribution of blood glucose levels to identify diabetes. Kumar & Ghosh (2024) emphasized the efficacy of BiLSTM-based deep learning models for the early identification of Parkinson’s Disease using online handwriting analysis. The suggested approach integrates sophisticated kinematic feature extraction with sequential modeling, providing a robust, efficient, and accurate solution that surpasses current machine learning methods. This method boosts diagnostic precision and offers a non-invasive, accessible tool for early diagnosis, improving patient outcomes.

Long short-term memory networks

RNNs that can manage long-term data dependencies include LSTM networks. LSTMs incorporate both short- and long-term memory by integrating a gating mechanism, as shown in Fig. 5, which shows how synchronized gating mechanisms in LSTM enable the selective retention and updating of diagnostic patterns across time, essential for understanding learning progression trends in chronic disease modeling. Mathematical notation is outlined in Eqs. (3)–(6).

Forget gate:

(3) $$f_t=\sigma (W_f\cdot [h_{t-1},x_t]+b_f).$$Input gate and candidate gate:

(4) $$i_t=\sigma (W_i\cdot [h_{t-1},x_t]+b_i),{\tilde{C}}_t=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}(W_c\cdot [h_{t-1},x_t]+b_C).$$Cell state update:

(5) $$C_t=f_t\cdot C_{t-1}+i_t\cdot {\tilde{C}}_t.$$Output gate:

(6) $$o_t=\sigma (W_o\cdot [h_{t-1},x_t]+b_o),h_t=o_t\cdot \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}(C_t).$$

Oktay & Kocer (2020) proposed a convolutional LSTM approach to distinguish between Parkinsonian Tremor (PT) and Essential Tremor (ET). They used a jump motion controller with a 4D camera to record tremors and tried to extract the key characteristics using a CNN. They classified the ET and PT using the LSTM network. However, the normalization method and strategy to prevent over-fitting were not disclosed. RNNs with LSTM are particularly effective at handling jobs involving sequential data. For instance, by investigating the temporal patterns of brain activity, LSTMs are used to detect cardiac disease.

A collection of heart rate variability measurements from heart disease patients and healthy controls was used to train an LSTM model. An accuracy of 85% was achieved in the model’s ability to recognize patients with heart disease. Goyal, Rani & Singh (2024) presented a multilayered framework based on deep learning that revolutionizes Alzheimer’s Disease diagnosis, attaining enhanced accuracy via transfer learning, LSTM-based temporal modeling, and GAN-facilitated data augmentation. The approach addresses critical issues in early detection, data scarcity, and overfitting, surpassing current methodologies and paving the way for future progress in personalized medicine and AI-driven diagnostics. Its prospective applications in early detection, multimodal integration, and clinical implementation underscore its importance as a crucial instrument in combating Alzheimer’s Disease.

Generative adversarial networks

GANs are a kind of deep learning model that produces artificial data. They are frequently employed in the healthcare industry to produce artificial medical images and supplement data. Figure 6 presents a GAN that generates synthetic patient data to supplement constrained clinical datasets, thereby improving the efficacy of deep learning in diagnosing uncommon diseases. They train antagonistic neural networks side by side. There are two networks: the Discriminator and the Generator and mathematical notation is outlined in Eqs. (7)–(9).

Generator:

(7) $$G(z;{\theta }_g)\to x^{\mathrm{\prime }}.$$Discriminator:

(8) $$D(x;{\theta }_d)\in [0,1].$$Minimax Objective:

(9) $$\underset{GD}{minmax}V(D,G)={\text{E}}_{x\sim pdata(x)}[\mathrm{l}\mathrm{o}\mathrm{g}D(x)]+{\text{E}}_{z\sim pz(z)}[\mathrm{l}\mathrm{o}\mathrm{g}(1-D(G(z)))].$$

The Generator creates bogus data samples to trick the Discriminator by increasing its chances of making a mistake. The Discriminator tries to discriminate between fake and actual samples. The two neural networks that build GANs are the generator and the discriminator networks. The discriminator network separates actual and artificial data, while the generator network creates artificial data.

Tufail et al. (2024) indicated substantial improvements in classification accuracy, sensitivity, and overall performance, underscoring the effectiveness of the suggested method. Using GANs to supplement minority classes improves the model’s generalizability and establishes a benchmark for using data augmentation in medical imaging. This study highlights the efficacy of integrating transfer learning with GAN-based augmentation to enhance the early identification of Alzheimer’s disease (AD), facilitating improved disease management and patient care.

Deep neural networks

An ANN with numerous layers of neurons is known as a DNN. The connections between the neurons in each layer and the subsequent layer are weighted. The connections’ weights are changed throughout the training phase so the DNN can learn how to do a particular job and mathematical notation is outlined in Eqs. (10), (11).

Layer wise transformation:

(10) $$h^{(l)}=f(W^{(l)}{h}^{(l-1)}+b^{(l)}).$$Final Output:

(11) $$y=f(W^{(L)}{h}^{(L-1)}+b^{(L)}).$$

Deep neural networks are being created for use in the healthcare industry by the research group Google DeepMind Health. Algorithms developed by DeepMind Health may be used to identify diseases, predict patient outcomes, and create customized treatment regimens (Singh, 2024). SinhaRoy & Sen (2024) improved the early diagnosis of Alzheimer’s Disease (AD) by producing synthetic Magnetic Resonance Imaging (MRI) images with Deep Convolutional Generative Adversarial Networks (DCGANs). The authors highlighted the efficacy of GAN-based methodologies in mitigating data constraints and improving diagnostic accuracy. It emphasizes an innovative method for Alzheimer’s disease prediction, providing a robust instrument for early diagnosis and enhancing the use of deep learning in medical imaging.

Generative artificial intelligence

Generative AI, which can produce new data such as text, images, or music based on training data, offers several fascinating potential uses in diagnosing and predicting diseases. It can uncover hidden patterns by analyzing vast patient data, including genetic information, scans, and medical records, and finding links and patterns humans might overlook. This capability can lead to earlier diagnoses, more precise prognoses, and even the identification of new diseases and mathematical notation is outlined in Eqs. (12), (13).

General generative model likelihood:

(12) $$p\theta (\mathrm{x})=\int p\theta (x|z)p(z)dz.$$Variational Autoencoder (VAE) objective (commonly used in GAI):

(13) $$\text{L}(\theta ,\varphi ;x)={\text{E}}_{q\varphi (z|x)}[\mathrm{l}\mathrm{o}\mathrm{g}p\theta (x|z)]-D_{KL}(q_{\varphi }(z|x)\Vert p(z)).$$

Additionally, generative AI can simulate disease progression by building models that mimic the course of a disease in a specific patient, assisting medical professionals in determining the optimal course of action and predicting a patient’s potential response to various treatments. Moreover, generative AI can produce synthetic data that resembles actual patient data, which is invaluable for developing new diagnostic tools and training other AI models (LeewayHertz & Takyar, 2024). Employing creative ideas to surpass conventional methods is crucial in healthcare. The use of GAI has been steadily rising across several domains. Balas & Micieli (2023) used generative artificial intelligence technology to produce visuals depicting the visual perception of a patient with visual snow syndrome, using textual descriptions to facilitate the text-to-image translation process. Diverse models provided a clear image, including DALL·E2, midjourney, and Stable Diffusion.

Overview of the dataset

Various chronic diseases and data modalities from public and commercial sources comprise the datasets used in existing works. Table 2 shows that the datasets used in medical research vary significantly in nature, availability, and accessibility, reflecting the varying data requirements of different disease areas.

Figure 7 illustrates a disease diagnostic pipeline based on deep learning, beginning with the data collection from medical imaging and clinical records. Preprocessing improves the data by normalizing, augmenting, and handling missing variables. Feature extraction employs deep learning algorithms.

The model undergoes training using data and is then precisely adjusted to enhance its accuracy. Disease classification predictions are assessed using performance metrics. Explainability approaches, such as Grad-CAM, emphasize significant characteristics, thereby improving the transparency of the model. The last stage involves incorporating the model into clinical operational processes, thus facilitating immediate diagnosis and bolstering healthcare decision-making. Figure 7 illustrates a comprehensive pipeline for illness detection using deep learning, highlighting the progressive integration of clinical data management, model development, and practical implementation. It distinctly integrates explainability and feedback mechanisms to enhance trust and decision-making in therapeutic environments.

Disease diagnosis and prediction

The study reviews numerous research studies that use DL methods to diagnose and forecast chronic diseases. These diseases include heart disease, cancer, diabetes, skin disorders, contagious diseases, and neurodegenerative diseases such as Alzheimer’s and Parkinson’s. The study focuses on advances in applying DL models and other deep learning architectures to improve diagnostic accuracy, early detection, and tailored treatment planning for various disorders. By evaluating varied datasets such as medical images, patient records, and genetic data, DL techniques have been demonstrated to considerably improve prediction results, allowing for quicker treatments and possibly lowering healthcare expenditures.

Algorithmic steps to implement deep learning models

The disease diagnosis and prediction process using deep learning encompasses many stages, including data collection, preprocessing, model selection, training, and deployment. This improves the precision of diagnoses and the ability to make predictions in medical environments. The algorithmic steps show the fundamental disease diagnosis and prediction stages for implementing deep learning models. This survey is a valuable guide for researchers and clinicians aiming to leverage deep learning in healthcare applications using algorithmic steps to design the DL model (Algorithm 1).

Model optimization and performance metrics

Hyperparameters are essential in influencing the performance, generalization, and efficiency of deep learning models in disease diagnosis. The Learning Rate (LR) is a crucial hyperparameter that regulates the magnitude of weight updates. A minimal learning rate (e.g., 0.001 for Adam or 0.01 for SGD) facilitates stable convergence and mitigates overshooting, whereas a greater rate accelerates learning but poses a danger of divergence. To maximize this parameter, researchers commonly apply tuning procedures such as grid search, cosine annealing, or cyclical learning rate scheduling. Another essential hyperparameter is the batch size, often ranging from 16 to 128 in medical imaging investigations. Reduced batch sizes enhance generalization due to increased stochasticity in gradient updates, whereas higher batch sizes optimize computational efficiency and use GPU parallelization. Researchers frequently determine batch size based on dataset dimensions and hardware limits, occasionally employing gradient accumulation when memory constraints inhibit bigger batches.

The number of epochs specifies the frequency with which a model traverses the dataset. The selection of epochs, typically ranging from 50 to 200 in medical deep learning research, is contingent upon the size and complexity of the dataset. Inadequate epochs may result in underfitting, while an excessive number of epochs poses a danger of overfitting. To mitigate this, strategies like early halting and monitoring validation loss are commonly utilized. The optimizer directly influences convergence. Adam, SGD with momentum, and RMSProp are commonly employed in disease diagnostic applications. Adam is renowned for its adjustable learning rate mechanism, which is effective on diverse medical datasets, although SGD offers more steady convergence, especially when utilized with momentum. Comparative testing is frequently used to identify the best appropriate optimizer.

Regularization methods, including dropout (rates of 0.2 to 0.5) and L1/L2 penalties (λ values from 0.0001 to 0.01), are employed to alleviate overfitting, a prevalent challenge in medical domains with insufficient data. Dropout randomly disables neurons during training, encouraging the network to acquire more resilient representations, whereas weight decay diminishes dependence on substantial weights. Cross-validation is frequently employed to optimize these parameters. The weight initialization approach is essential. Xavier (Glorot) and the initialization are commonly utilized to stabilize gradient propagation in CNNs, GANs, and DNNs, hence mitigating the hazards of disappearing or bursting gradients. Although they are frequently established by architecture, they substantially influence initial training dynamics.

In non-linear representation learning, activation functions like Rectified Linear Unit (ReLU) and Leaky ReLU are frequently employed in CNNs and RNNs, whereas Gaussian Error Linear Unit (GELU) and Swish have gained prominence in Transformer-based architectures like ViTs because of their smoother activation characteristics. The output layer often uses softmax for classification or sigmoid for binary jobs. In vision-based illness diagnosis, the size of the input picture is a significant hyperparameter. Standard dimensions, like 224 × 224 (ResNet, ViT) and 299 × 299 (Inception), are employed to reconcile spatial detail with computing efficiency. Larger medical images, such as 512 × 512 MRI or CT scans, are frequently shrunk or cropped according to model capability. In Vision Transformers, patch size (often 16 × 16 or 32 × 32) is a critical hyperparameter; smaller patches retain intricate features of lesions, whilst bigger patches decrease processing demands.

Learning rate schedulers, including step decay, cosine annealing, and ReduceLROnPlateau, are frequently utilized to dynamically modify the learning rate throughout the training process. These strategies promote convergence, mitigate stagnation, and improve generalization, especially in multimodal and unbalanced datasets. These hyperparameters together impact predicted accuracy, training efficiency, robustness, and clinical application. Meticulous selection, informed by empirical tuning methodologies such as cross-validation, Bayesian optimization, or adaptive scheduling, is crucial to enhancing the efficacy of deep learning models in disease detection.

Gupta et al. (2025) employed fivefold cross-validation (performed five times) on a coronary CT dataset, documenting average accuracy and AUC throughout the folds. The authors specifically conducted fivefold cross-validation five times to stabilize their results. Kong et al. (2025) employed 5-fold cross-validation to train deep MRI classifiers for the selection of appropriate image inputs. Formal hypothesis testing is also utilized. The glioma/MRI study employed t-tests and Mann–Whitney U-tests to compare patient groups, whereas other studies utilized paired t-tests or bootstrap tests on cross-validated ratings for comparing model variants. Huang et al. (2023) Performance indicators (accuracy, AUC, F1, etc.) are invariably accompanied by uncertainty estimates, typically represented as 95% confidence intervals. The CT model attained an AUC of around 0.95 with a 95% confidence interval of 0.92–0.97, while a transformer-based EHR model demonstrated an AUROC of 0.844 (95% CI [0.838–0.851]) on the internal test set and 0.849 (95% CI [0.846–0.851]) on an independent cohort.

Importantly, the majority of studies incorporate external validation using independent datasets. Models are evaluated on reserved cohorts or distinct hospital datasets to exhibit generalization and robustness. To identify the location of kidney stones, Ma et al. (2020) proposed a Heterogeneous Modified Artificial Neural Network (HMANN) approach that lowers noise and aids in kidney image segmentation. Three classifiers were tested to arrive at the correct prediction results for Chronic Kidney Disease (CKD), multilayer perceptron, artificial neural networks, and support vector machines and AUC, while Gupta et al. (2025) attained consistent AUCs with narrow confidence intervals on an unobserved cardiac dataset. Authors frequently observe that external testing “exhibits robustness and generalizability.” Certain research enhance reliability by conducting repeated experiments (e.g., multiple cross-validation runs) or use bootstrapping techniques. Gupta et al. (2025) specifically conducted fivefold cross-validation five times to stabilize their results. Statistical significance tests, external test-set performance, 95% confidence intervals for AUC/accuracy, and k-fold cross-validation data are all used by contemporary deep-learning diagnostic investigations to prove that their findings are reliable and repeatable.

Table 3 presents a comprehensive summary of the performance evaluation measures often utilized in deep learning-based illness detection. The table delineates classification and regression metrics, as well as segmentation and calibration measures, providing a comprehensive overview of model assessment in medical contexts. In classification tasks, accuracy, precision, recall (also known as sensitivity), specificity, F1-score, AUC-ROC, and AUC-PR are commonly used to assess correctness, sensitivity in disease detection, and threshold-independent separability of predictions. Regression measures such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the R² score are employed in scenarios like illness progression tracking or severity grading. Image-based applications, such as tumor and lesion detection, depend on spatial overlap metrics like the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), which quantitatively assess the concordance between predicted and actual anatomical areas.

Furthermore, probabilistic metrics such as the Brier Score and calibration curves assess the dependability of anticipated probabilities, hence enhancing confidence in clinical decision-making. By summarizing the formulae, clinical significance, and popular use cases, the table not only gives clarity on metric selection but also stresses the need of employing many complementing measures for a full and clinically useful evaluation of deep learning models.

Large dataset

Deep learning models need substantial computational resources. A basic deep learning model requires more network parameters in its architecture. The efficacy of deep learning models is significantly contingent upon the size of the training dataset. Due to their automated feature extraction capability, deep learning models need exposure to several variations of instances within the corresponding classes in the dataset to achieve effective generalization. This explains the success of deep learning in domains where large quantities of data can be readily gathered. A substantial volume of data may be amassed across several areas, such as natural language processing and computer vision, resulting in the increasing popularity of deep learning applications in these fields. Nonetheless, producing vast quantities of data poses a considerable challenge in medicine. Consequently, we cannot adequately train deep learning models using limited datasets to attain the intended objectives. Moreover, medical data analysis is more complex than other fields, such as image or voice recognition. These complexities impair the model’s capacity for effective generalization. A substantial amount of training data is hence necessary to mitigate this problem. Innovative data generation methodologies must enhance the volume of data collected and utilized for dataset formulation in the medical domain (Lopez Alcaraz et al., 2025).

Deep learning methods need extensive and varied datasets for successful generalization, since exposure to various examples aids in the acquisition of strong representations of fundamental patterns. In medical fields, databases frequently exhibit limited sizes and imbalances due to difficulties in acquiring annotated patient information. Insufficient data heightens the likelihood of overfitting, causing models to remember training samples instead of identifying genuine disease-related characteristics, which leads to diminished test performance and skewed predictions for underrepresented populations. The absence of generality is especially alarming in healthcare, since it may result in erroneous diagnosis or misclassification of uncommon illnesses. Research indicates that limited datasets “fail to generalize patterns,” resulting in unstable decision limits and diminished dependability when utilized with unfamiliar patient data.

To solve these issues, researchers apply methodologies such as transfer learning, regularization, ensemble models, and data augmentation. Data augmentation is very efficacious in improving model performance when working with limited datasets. Through artificial dataset expansion, augmentation introduces increased variability to models while maintaining diagnostic labels. Prevalent methodologies encompass geometric and photometric alterations (e.g., rotation, flipping, brightness modifications) for medical imaging, as well as Synthetic Minority Over-sampling Technique (SMOTE) or Conditional Tabular Generative Adversarial Network (CTGAN) for tabular clinical datasets. Generative models, such as GANs, are employed in advanced techniques to generate realistic MRI or X-ray images, which have been shown to improve the accuracy of Alzheimer’s and cancer identification. Data augmentation is a critical tool for the development of dependable deep learning systems in healthcare contexts with limited data, as it mitigates overfitting, improves generalization, and equilibrates class distributions.

Availability of public and real-time disease diagnosis datasets and ethical considerations

There are publicly accessible databases for various diseases; however, collecting or obtaining real-time data remains challenging, owing to ethical concerns. Patient permission is required before data collection to guarantee privacy and compliance with ethical guidelines. Real-time medical data can only be released as open source after successfully deploying the DL models, protecting patient anonymity. Access to public and real-time disease diagnosis databases is essential for improving healthcare research, especially in creating AI-driven diagnostic systems. Many datasets exist across many disease areas, each providing distinct insights into certain diseases. Access to these datasets often requires compliance with ethical principles and data-sharing agreements to safeguard patient privacy and data security. These resources provide a basis for pioneering research in disease diagnosis, enabling the creation of more precise, real-time diagnostic instruments that may eventually enhance patient outcomes (Borda et al., 2022).

Quality and availability of data

Deep learning models demand substantial quantities of labeled training data for efficient learning. Acquiring enough high-quality labeled data may be costly, time-intensive, or complex, especially in specialist fields or when handling sensitive information such as healthcare and cybersecurity. Despite several methods, like data augmentation, to produce substantial data, generating sufficient training data to meet the demands of deep learning models may sometimes be arduous. Moreover, a limited dataset might result in overfitting, causing deep learning models to excel on training data but struggle to generalize to novel data. Balancing model complexity and regularization methods to prevent overfitting while attaining effective generalization is difficult in deep learning. Furthermore, investigating methodologies to enhance data efficiency, such as few-shot, active, or semi-supervised learning, remains a prominent study domain (Talaei Khoei, Ould Slimane & Kaabouch, 2023).

Data enrichment methods and the time complexity of the model

Employing conventional data augmentation techniques facilitates the introduction of significant variances in data samples for training. Nonetheless, the quality of the enhanced instances deteriorates beyond a certain threshold of augmentation. Consequently, novel data enrichment procedures are necessary for disease diagnostic tools. Generative Adversarial Networks (GANs) are used throughout several fields, including medical image synthesis, to provide high-quality data samples. Additional strategies are necessary to synthesize high-quality data. Deep learning models for medical diagnoses must manage large amounts of data, leading to increased network complexity. This requires additional processing resources and raises the model’s computational complexity. To address model complexity, deep learning models need high-end computers with sufficient CPUs and GPUs (Shyamala & Navamani, 2024).

Lack of cohesive frameworks and suboptimal utilization of transformer architectures

Existing research works concentrated on attaining high accuracy within certain domains (e.g., lung cancer or heart disease) using single-modality data and model types. Nevertheless, they lacked a cohesive framework that amalgamated essential components such as interpretability (XAI), privacy-preserving techniques, Transformer-based topologies, and cross-modal data fusion. This fragmentation hinders the development of resilient, practically deployable systems. Despite their shown efficacy in sequential biological data, Transformers have been little investigated. Research works partly integrate attention processes; nevertheless, no research has comprehensively used Transformer models for EEG, ECG, or text-based features. This indicates a lost chance to represent intricate temporal and contextual linkages in clinical data (Shatnawi, Abuein & Al-Quraan, 2025).

Inadequate hyperparameter tuning and uneven datasets

Numerous research studies failed to provide detailed documentation of tuning methodologies. While a few research works exhibited performance enhancements via systematic tuning, others depended on default configurations, compromising repeatability and possible performance advancements. Data imbalance was a considerable challenge, particularly in the lung cancer and dermatological datasets. While CTGAN and SMOTE were used in some instances to mitigate this issue, most research failed to implement thorough class-balancing procedures or efficiently utilize data augmentation. This raises problems over the model’s generalizability to diverse real-world populations (Alzahrani, 2025).

Absence of privacy-enhancing techniques and limited external validation and generalizability

Only a few research works have investigated federated learning or blockchain as methods to facilitate collaborative learning while safeguarding patient data privacy. Considering the sensitivity of medical data, this is a significant oversight that must be rectified in further study. External validation was either absent or inadequately highlighted in the majority of research. While research works validated across many datasets, most mainly depended on single-dataset, training-testing split assessments, limiting confidence in the models’ efficacy on unobserved populations and diverse clinical environments (Abbas et al., 2025).

Transparency

Deep learning’s interpretability and explainability are significant roadblocks to deciphering complicated models. The decision-making procedures of increasingly complex deep neural networks, with their many layers and parameters, might seem like “black boxes,” making it hard to understand the reasoning behind individual predictions. The lack of transparency in areas with high stakes, such as healthcare and finance, makes it harder to trust and use these models. Finding an acceptable balance between model performance and readability is essential for researchers, regulators, and end-users to understand the model’s logic, which is necessary for making informed decisions and holding models accountable in the complex world of contemporary deep learning (Hadadi & Arabani, 2024).

The present investigation exhibits encouraging outcomes; nonetheless, it is constrained by many limitations. The dataset used was constrained in size and variety owing to the availability of annotated data, potentially impacting the model’s generalizability to wider populations. Secondly, although performance was assessed using standard measures, external validation using unobserved clinical datasets was not conducted because of limitations in data availability. Third, computing resource limits hindered the deployment of more complicated Transformer-based structures and significant hyperparameter tuning. Additionally, privacy-enhancing approaches such as federated learning or differential privacy were not implemented in this model version but are planned for future work to ensure compliance with data confidentiality regulations. Despite these difficulties, deep learning can potentially enhance disease diagnosis and prognosis. Large quantities of patient data can be analyzed using deep learning to find patterns that are invisible to human clinicians. This can help diagnose people at risk of acquiring specific diseases and result in more precise and prompt diagnoses.