Skin diseases represent a prevalent category of conditions that affect people of all ages globally, posing considerable challenges for healthcare managementr1_alruwaili2025integrated. The WHO published a report where nearly 900 million individuals worldly suffering different skin diseases, with the most common being onychogryphosis, acne, atopic dermatitis, eczema, psoriasis, rosacea, cyanosis, clubbing, and koilonychiar2_hay2014global. Fungal infections are the most common skin condition, affecting 6.3 per 1,000 people, followed by psoriasis (3.7), urticaria (2.9), atopic dermatitis (1.3), and vitiligo (1.3 per 1,000 population)r3_kavita2024self. According to the report, about 10% of all dermatological conditions primarily affect older individuals adultsr4_maddy2018hair. The general health status of a person is expressed by nails as an important indicator of good health. In general, pink nails indicate good health, and healthy nails appear smooth and have a uniform colorr5_wollina2016diagnosis. Additionally, nails become weak due to a lack of oxygen supply from the body. Signs of illness are also demonstrated simultaneously through the nailsr6_nithya2019nail. Currently, doctors identify most skin diseases when the symptoms of the disease are correctly identified. Today, dermatologists use autonomous disease diagnosis systems to reduce the use of medications. In addition, computer-assisted diagnosis provides inexperienced clinicians with the ability to identify difficult casesr7_wollina2016diagnosis. Recent advances in deep learning, particularly convolutional neural networks (CNNs), have greatly enhanced medical image diagnosis with high automation and accuracy. Pre-trained CNN models like EfficientNetr8_ebadi2025abnormality, ResNetr9_nasra2025resnet18 and DenseNetr10_ramadevi2025deepnet consistently deliver optimal performance in classifying skin lesions. Most models use transfer learning to extract relevant features, enabling precise medical image predictions without the time-consuming need to annotate large datasets. Our developed model addresses the present key gaps in skin disease identification. the hybrid model are combining multiple pre-trained architectures and the hybrid model leverages their complementary strengths to enhance performance. Moreover, batch normalization and dropout techniques are used to mitigate overfitting and manage class imbalance.

This research has made several key advances:

The use of adanced technology in medical field are increased grudually especially deep learning techniques are very effective for classifying skin diseases. Previous research in assistive technology has focused on integrating sensors with IoT r11_hassain2025smart, r12_hassain2025voice, r13_hassain2024iot, r14_huria2024iot, r15_kader2024head, r16_kader2024wireless, r17_hassain2023design and applying computer vision-based image analysisr18_kader2025sign,r19_nesa2025image to advance medical care.

In r20_nijhawan2017integrated the authors developed a CNN-based deep learning model for nail disease identification, achieving a system accuracy of 84.58%. The authors in r21_muhamad2019application developed a CNN-based system for the early detection of Terr’s Nail abnormalities, and results demonstrated an accuracy of 95.24%. The authors in r22_rahman2020transfer developed a leukonychia nail disease detection system using the deep learning model DenseNet201. It compared five transfer learning models and selected three of the most dangerous classes of nail diseases for detection, achieving an accuracy of 93.8%. In r23_nijhawan2017integrated the authors developed an CNN-based model to identify 11 types of nail diseases using a hybrid of CNNs, achieving an accuracy of 84.58%. The authors in r24_hamim2023multiple integrated image processing methods with models, including MobileNetV2, VGG16, and VGG19, to classify nail diseases, the MobileNetV2 model obtains the maximum accuracy of 83%. The authors in r25_yilmaz2021deep developed a nail disease detection model using deep neural networks with VGG16 and InceptionV3, where VGG16 achieving a highest accuracy rate of 95.98% in detecting fungal infections. In r26_sougukkuyu2023classification, the author developed an AI image analysis system using a transfer learning-based VGGNet deep CNN architecture. The VGG-16 model was tested for early diagnosis of nail diseases, achieving an accuracy of 94%. In r27_regin2022nail, the authors proposed an ensemble of CNN-based model to detect and categorize nail disorders using photographic data, achieving an accuracy rate of 95%. According to a research r28_begum2021automated, developed a CNN-based automated system, which used “DermaDoc” web app, to identify nail syndromes and related skin diseases. By using data augmentation and transfer learning, the model achieved 92.5% accuracy. In r29_marulkar2023nail, the author identified nail illnesses using a CNN deep learning model, achieving an accuracy of 87.33%. The authors in r30_abdulhadi2021human developed a human nail disease classification system using transfer learning with CNN models, where DenseNet201 obtain accuracy 96.39%.

Moreover, another system was developed by the authors to detecting various types of fingernail illnesses using the EfficientNet-B2 model, achieving highest accuracy of 72% r31_can2022diagnosing. The authors developed a nail disease classification system combining deep learning with graph-based learning using Graph Attention Network and ResNet, achieving an accuracy of 87.91% r32_indriani2025nail. The authors in proposed nail disease detection using the ResNet18 architecture. The model achieved an accuracy of 91% to detect various nail diseasesr33_nasra2025resnet18. A automated system for classifying nail diseases using CNN-based DenseNet121 was introduced by the authors in. The model achieved an accuracy of 81.3% and accurately nail disease detection through deep learningr34_kaur2025smart. This research introduces a hybrid model to classify six types of nail diseases, using VGG16 to extract image features and SVM for classification. It can achieve 94.36% accuracyr35_singh2025hybrid. In r36_macriga2025medical, the authors introduced an automated system for detecting and classifying nail diseases using CNN. Analyzing a large dataset of nail images, the system achieves accuracy 92%. The authors in r37_pakpahan2025integration developed a web-based computer vision predictive early detection system for nail diseases by YOLOv8 object detection model with the FastAPI architecture framework. The model real-time nail disease identification with four-class and achieves an accuracy of 93%. The research uses the YOLO-NAS algorithm to identify and classify skin diseases from a diverse dataset of conditions including cellulitis, impetigo, athlete’s foot, nail fungus. With data preprocessing the system achieves 78% accuracyr38_deepa2025deep. A hybrid nail diseases detection system was developed by the authors in r39_roy2024emperical, where CNN-LSTM model for classifying various nail conditions. The model performs especially well detecting nail fungus and achieved accuracy of 94%. The authors developed a nail disease recognition framework using the DenseNet121 model to analyze and classify six types of nail diseases from images, achieving 86% accuracyr40_kaushik2024enhanced. To detecting nail diseases the research uses the DenseNet-121 model to classify six nail conditions, achieving 84.6% accuracyr41_chuenchit2024classification. The research presents a hybrid CNN-Random Forest model for classifying four dermatological conditions, achieving 96.08% accuracy to disease detection accuratelyr42_kumar2024disease. The authors in r43_singh2024deepfungusdet developed a deep learning approach using MobileNetV3 to detect fungal infections. The model achieved 93.14% accuracy in identifying various fungal conditions. The research detects nail diseases though Mask R-CNN by analyzing nail image. It also predicts underlying systemic diseases and the system achieves an accuracy of 82% in disease and severity identificationr44_prajeeth2023smart.

The model was trained using the Adam optimizer with a learning rate of 0.0005. The training was conducted with a batch size of 16 and for 15 epochs, incorporating an EarlyStopping callback with a patience value of 2 and to prevent overfitting and retain the optimal model state. Mixed-precision training was enabled to accelerate computation and reduce memory usage on compatible NVIDIA GPUs with Tensor Cores, while ensuring that the final output layers maintained precision for numerical stability. A ModelCheckpoint callback was used to automatically save the best-performing weights during training. The batch size of 16 is optimal for Colab GPU environments with 128×128 input images, though mixed precision can permit larger batch sizes depending on available VRAM.

In the Figure 4, the training and evaluation accomplishment curves are demonstrated. The graph of training and validation accuracy over epochs shows rapid improvement from the initial epoch, with both metrics approaching near-perfect levels by the sixth epoch. Training accuracy rises from 95.7% to 99.7%, while validation accuracy starts at 0.993 and remains stable thereafter, finally reaching approximately 100%, indicates that the model avoids overfitting, pointing to balanced data, suitable regularization, and proper model complexity. In Figure 5, the training loss begins at approximately 0.10, indicating a moderate error rate as the model starts learning feature representations from the dataset. The hybrid model converges quickly after the loss decreases sharply within the first few epochs and training loss levels off at around 0.002 in the third epoch. The hybrid model learns from the steady decline in both losses shows that the training data performs well on new, unseen data, avoiding overfitting and it can highlight the model's ability to reduce prediction errors across different data sets and it can set close values of the training and validation losses, which are near zero. The Curve of validation loss follows a same valuation confirming that the model is not follow overfitting or underfitting. Validation Loss Performance Curves starts from lower and it can stays simultaneously low throughout training without any change.

In Table 1, the Performance Metrics of Nail Disease Detection for different models are shown. The accuracies achieved were: LinearSVM 97.86%, KNN 97.53%, Light CNN 97.52%, MobileNetV2 95.40%, Random Forest 84.13%, Decision Tree 76.46%, and EfficientNetB0 73.66%, while the hybrid ensemble model reached an accuracy of 98.33%. The highest accuracy was obtained by the hybrid model among the evaluated models. The Hybrid model achieves the highest precision at 98.89% that can indicates the proportion of predicted positive cases that are actual positives. The deep learning models Linear SVM, KNN, and Light CNN are nearly follow closely behind. In the other hand EfficientNetB0 model has the lowest precision rate 74.00%. Hybrid model Recall identifies true positives cases is highest with results 98.83%. Other models like Linear SVM, KNN, and Light CNN have approximately similar sensitivity, while the EfficientNetB0 model falls behind with a lower recall of 73.83%. The Hybrid model shown that the highest F1 Score is 98.83%. The result highlighting that model have excellent balance between correctly identifying true positives with minimizing false positives. In the case of EfficientNetB0 and Decision Tree models demonostrate it nearly lower F1 Scores in between 73% to76% and output are indicating less due to imbalances between precision and recall classification performance.

In Figure 6, we demonstrate the comparison of the outputs of eight deep learning models. The Hybrid Ensemble model gets the best results on all metrics, with scores around 0.99, which is better than any of the other single models. While Random Forest, Decision Tree, and EfficientNetB0 demonstrated comparatively poorer performance, with scores ranging from 0.70 to 0.85, Linear SVM, KNN, and Light CNN also shown strong performance, with scores between 0.95 and 0.98. This finding indicates that the suggested hybrid model successfully combines the positive aspects of several algorithms, providing improved nail disease detection accuracy, stability, and dependability. The reliability is higher with the Hybrid Ensemble technique than with single models. When compared to single models, the Hybrid Ensemble technique exhibits higher reliability. With an accuracy of 99.70%, the hybrid model is very dependable for the practical diagnosis of nail disorders. The Hybrid model leads with 98.89% precision, closely followed by Linear SVM, KNN, and Light CNN, while EfficientNetB0 trails with 74.00%. Precision is a measure of the accuracy of positive predictions. In contrast to EfficientNetB0, which scores much lower at 73.83%, the Hybrid model achieves a peak recall of 98.83%, with Linear SVM, KNN, and Light CNN also demonstrating strong performance. EfficientNetB0 and Decision Tree models have substantially lower scores of roughly 73-76%, whereas the Hybrid model's robustness is highlighted by its 98.83% F1 Score, which balances precision and recall. EfficientNetB0 and Decision Tree models have substantially lower scores of roughly 73-76%, whereas the Hybrid model's robustness is highlighted by its 98.83% F1 Score, which balances precision and recall. This demonstrates unequivocally how well the Hybrid ensemble can classify nail diseases in a balanced and trustworthy manner.

Six models (KNN, Light CNN, MobileNetV2, Random Forest, Decision Tree, and Linear SVM) used to classify nail diseases are shown in Figure 7 alongside with their confusion matrices, which are divided into two categories: "Healthy Nail" and "Nail Fungus Disease." On the other hand, the model only misclassified 9 cases as healthy while accurately identifying 741 cases of sickness. The KNN model's low rates of false negatives (missed disease cases) and false positives (healthy nails mistakenly flagged as diseased) highlight the model's dependability for clinical screening applications. These results correlate to high recall values for both classes—0.96 for healthy nails and 0.99 for diseased nails. In Figure 7(b),shows the confusion matrix for the Light CNN and the model correctly identified742 healthy nail cases are True Negatives with a 99% recall and 721 nail funguscases are True Positives with a 96% recall. It can incorrectly identify 29 disease cases as healthy while misclassifying only 8 healthy nails as diseased. The model's ability to detect nail fungus with high precision and strong sensitivity without bombarding clinicians with false alarms is demonstrated by its low error rate. As shown in Figure 7(c), the MobileNetV2 model correctly classified 98% of diseased nails, achieving a high true positive rate. Additionally, 94% of healthy nails were accurately identified. Nevertheless, compared to other models, a slightly higher false positive rate (6%) indicates that more healthy nails were mistakenly identified as diseased. Very few diseased cases are missed by the model, as evidenced by the low false negative rate of 2%. In Figure 7(d), the Random Forest model's performance is shown. Out of 750 actual healthy nails, it correctly identified 554 as healthy, resulting in a high True Negative rate of 74%. However, it also misclassified 196 healthy nails as having the disease, meaning 26% of the healthy cases were False Positives. For nail fungus disease, the model achieved a 94% True Positive rate, meaning it correctly identified most diseased cases, while 6% were False Negatives.

In Figure 7(e), the Decision Tree model demonstrates a strong ability to detect nail fungus disease, correctly identifying 91% of true cases and it misclassified 284 healthy nails as having the disease, leading to a high False Positive rate of 38%. In Figure 7(f), shown the LinearSVM model demonstrates exceptional performance in identifying disease nails cases correctly classifying 97%. In Figure 8(a), illustrates the confusion matrix of the EfficientNetB0 model, which correctly identifies 71% of healthy nails and 77% of nail fungus cases, while misclassifying 29% of healthy nails as diseased and missing 23% of actual infections. In Figure 8(b), the Hybrid Ensemble model demonstrates strong performance. The model shows a high number of correct predictions for Normal Nails (450), with only 15 misclassifications. For the Normal Nail Disease class, it correctly identified 435 cases and recorded zero misclassifications, demonstrating perfect precision for disease detection. In Figure 9, the ROC curve analysis compares the performance of four models: KNN, Light CNN, MobileNetV2, and Random Forest. In Figure 9 (a) and Figure 9 (b), both the KNN and Light CNN models achieve perfect ROC curves with an area under the curve (AUC) of 1.00. Figure 9 (c), are the MobileNetV2 model also performs excellently, with an AUC very close to 1.00 (0.99), showing high true positive rates with minimal false positives. The Figure 9 (d), shown Random Forest model has a lower AUC of 0.93, suggesting relatively lower but still strong classification performance compared to the other models. Overall, KNN and Light CNN demonstrate the highest effectiveness, followed closely by MobileNetV2, while Random Forest shows good. In Figure 10 (a) and (b), provides a ROC curve analysis of the Decision Tree and EfficientNetB0. Interestingly, despite the typical strength of deep learning architectures, the Decision Tree slightly surpasses EfficientNetB0 in performance, as reflected by its higher AUC score of 0.82 compared to 0.79. However, the experimental results provide a crucial insight, where model effectiveness is context-dependent, and simpler models can occasionally outperform more complex ones when they are better aligned with the nature of the dataset and the problem at hand.

In Figure 11, the comparison with different models is demonstrated. the accuracy of eight models—Linear SVM, KNN, Light CNN, MobileNetV2, Random Forest, Decision Tree, EfficientNetB0, and the proposed Ensemble (Hybrid). The Ensemble (Hybrid) model achieved the highest accuracy nearly 100%, outperforming all others. Linear SVM, KNN, and CNN also performed well between 97 to 98%, while Random Forest, Decision Tree, and EfficientNetB0 showed lower accuracy between 75 to 85%. These findings demonstrate that the hybrid model produces predictions for nail disease detection that are more robust and dependable.

EfficientNetB0 performs worse than it should due to a number of factors, including a small dataset size, a relatively low image resolution (128 x 128 pixels), and possibly inadequate fine-tuning. To fully utilise its capabilities, its intricate architecture necessitates extensive data augmentation and significant hyperparameter optimisation. Another factor is that applying pretrained weights from generic image datasets to specialised medical images, such as nails, causes a domain shift because the learnt features might not best capture the subtle colour and textural variations that are essential for differentiating nail diseases, which lowers the discriminative power.The hybrid ensemble model, on the other hand, performs exceptionally well by fusing the various advantages of several classifiers. It effectively reduces bias and variance by compensating for the individual flaws of each model by combining their predictions. Through cooperation, overall accuracy and robustness are improved, resulting in more dependable performance under a range of data conditions. For medical image analysis, the synergy from this ensemble approach makes use of the complementary strengths of transfer learning and classical machine learning models, producing better generalisation and consistent classification results.

A comparison of earlier studies on different machine learning models for disease diagnosis is shown in Table 2. The reviewed literature, which covers the years 2017–2025, shows that current models, such as DenseNet, VGG16, MobileNet, ResNet, and various CNN architectures, achieve accuracies between 72% and 96.39%, with multiple studies reporting results in the 92–96\range. Although many studies lack transparency regarding dataset size, it is noteworthy that dataset sizes vary greatly, ranging from as few as 185 images to 15,000 images. In addition to using the largest dataset (15,618 images), the suggested hybrid ensemble model, which combines multiple architectures, surpasses all previous works with an outstanding accuracy of 99.70%. This outcome shows that the ensemble approach successfully utilises the strengths of constituent models, as it represents a significant improvement over the previous highest accuracy of 96.39% (DenseNet201, 2021) and 96.08% (Hybrid CNN-RF, 2024). This finding and the recent shift towards hybrid approaches point to a distinct trend in the field: combining several models provides a potent method for improving diagnostic accuracy in the classification of nail disorders.

Our findings show that the hybrid model successfully reduces individual flaws while generating a strong and trustworthy prediction system. Clinically, this ensemble approach shows promise for practical implementation, helping dermatologists diagnose nail diseases automatically and accurately, especially in rural or low-resource healthcare settings where access to experts is limited. Further improving accessibility through integration with cloud-based or mobile platforms could make teledermatology and remote diagnosis possible. Reliance on a single-center dataset is one of the limitations. To ensure transparency and foster clinician trust, future research should concentrate on integrating explainable AI (XAI) techniques, extending to multi-center datasets, and implementing mobile or portable diagnostic solutions for wider accessibility. To sum up, the hybrid ensemble model offers remarkable clinical integration potential, providing quick and precise nail disease classification, which helps with early diagnosis and better patient care.