http://dspace.dtu.ac.in:8080/jspui/handle/123456789/47 2026-10-26T15:03:39Z 2026-10-26T15:03:39Z ARORA, PRIYANKA Dahiya, Sonika (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23406 2026-10-16T06:38:28Z 2026-10-01T00:00:00Z

Title: ASSESSMENT OF IMPACT OF COVID-19 ON MENTAL HEALTH USING MACHINE LEARNING TECHNIQUES Authors: ARORA, PRIYANKA; Dahiya, Sonika (SUPERVISOR) Abstract: Major mental health disorders, like depression and anxiety, constitute a growing global public health challenge grossly exacerbated by large-scale crises, such as that presented by the COVID-19 pandemic. The unprecedented disruption brought on by prolonged lockdowns, social isolation, economic uncertainty, and health-related stressors has increased psychological distress throughout diverse populations, leading to an increase in undiagnosed and untreated mental health disorders. The early recognition and correct classification of depression and anxiety are thus very important to facilitate timely intervention and the development of appropriate mental health policies. Against this background, Machine Learning (ML) and Deep Learning (DL) approaches have emerged to have potential for use in the automated assessment of mental health. The proposed thesis seeks to move the state-of-the-art in automated detection of depression and anxiety by thoroughly examining classical ML, DL, and hybrid intelligent techniques based on psychological assessment data obtained during and post the COVID-19 era. Contrary to several prior approaches that have a narrow focus and binary classification boundaries, this study will prioritize scalability and multiclass prediction tasks. The thesis starts by performing a systematic literature review on ML,DL based mental health evaluation techniques, particularly in terms of COVID-19-related depression and anxiety. The literature survey analyzes different techniques, datasets, instruments, feature designs, and metrics that are currently used. Critical gaps in the literature are identified, including the absence of long-term studies, multi-class studies, the consideration of the size of datasets in analyses, rigorous statistical validation of the models, and the use of learning framework combinations. The thesis also examines post-pandemic mental health outcomes, highlighting emerging trends and research vi directions. These areas provide a foundation to the methodological contributions made in this thesis. Inthesubsequentphase,thethesisexaminestheeffectofthesizeofthedatasetonthe predictive capabilities of the traditional classifiers based on ML, such as the K-Nearest Neighbor (KNN) classifier, the Decision Trees (DT) classifier, the Random Forest (RF) classifier, and the Support Vector Machine (SVM) classifier. The experiment will make use of mental health survey data obtained during the COVID-19 pandemic. The results show that the model performs better with an increase in the size of the dataset, and it helps in understanding optimal levels of dataset collection needed to make a mental health model. The work is further continued with a comparative assessment of DL models to classify depression and anxiety. Different DL models like Artificial Neural Network (ANN), Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), Convolutional Neural Network (CNN), and TabNet are used on three different mental health datasets. The data shows that the performance of DL models,especiallyLSTMandCNN,issignificantlybetterthantraditionalMLmodelsin modeling complex interactions in features and multiclass classification. This outcome points to the effectiveness of applying DL models to model psychological patterns related to depression and anxiety. However, to circumvent some limitations in single models, some models in this research use a combination of several models with complementary properties. This combination will be in the form of hybrid classifiers; some examples include DT+KNN classification and RF+NNclassification. Theempiricaloutcomeshaveclearlyindicated the ability of the proposed hybrid models to provide significantly higher levels of predictive accuracy and robustness than the stand-alone classifiers, performing up to 99-100% accuracy on some datasets. This confirms the effectiveness of the hybrid learning approach for precise mental health prediction. The originalities in the thesis lie in the formulation of two predictive models for the detection of depression and anxiety. Firstly, the development of the multiclass Depression Anxiety-TabSVM (DA-TabSVM) hybrid model has taken place in order to tackle the complex, nonlinear nature of the data in the psychological tables. Secondly, the formulation of the Adaptive Sigma KNN (ASKNN) model has taken place. The vii model has the unique ability to alter the weight given to the surrounding samples through the adaptation of the sigma parameter. In traditional KNN models, this has not been the case. Thorough experimental studies, including cross-validation, analysis of Area Under the Curve– Receiver Operating Characteristic (AUC-ROC) curves, and statistical significance tests with the Friedman and Wilcoxon signed-rank tests, support that the new models clearly outperform existing baselines in ML. The effect-size analysis via Cohen’s d and Cliff’s Delta values strengthens the relevance of the obtained improvements. Furthermore, this thesis proposes a Feature-Weighted Adaptive Kernel KNN(FW-AK-KNN)modelthatincorporates feature importance and adaptive kernel basedneighborweighting toenhancebothaccuracyand interpretability inmental health prediction. On thewhole, this thesis makes a valuable contribution to a complete, scalable, and statistically valid framework for the automatic diagnosis of depression and anxiety using ML,DL,andhybridmodels. Withthedevelopmentoftrustworthyandefficientmethods of identifying mental health disorders at an earlier stage, the presented solutions contribute to global initiatives with regard to mental health issues and are in line with the United Nations (UN)’s Sustainable Development Goal (SDG) 3 and the importance of AI solutions in managing this burden worldwide.

2026-10-01T00:00:00Z KUMARI, ASHISH Kumar, Shailender (SUPERVISOR) Raw, Ram Shringar (CO-SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23403 2026-10-16T06:36:44Z 2026-07-01T00:00:00Z

Title: PERFORMANCE ENHANCEMENT OF VEHICULAR AD HOC NETWORKS USING INTELLIGENT TECHNIQUES Authors: KUMARI, ASHISH; Kumar, Shailender (SUPERVISOR); Raw, Ram Shringar (CO-SUPERVISOR) Abstract: Vehicular Ad Hoc Networks (VANETs) form the communication backbone of Intelligent Transportation Systems (ITS), enabling vehicle-to-vehicle and vehicle-to infrastructure exchange for safety-critical and traffic-management services. However, the high mobility of vehicles, rapidly changing topology, fluctuating density, and intermittent connectivity make stable and reliable communication difficult, producing unstable links, frequent cluster reformation, control overhead, and degraded performance. Clustering mitigates these problems by organising vehicles into manageable groups, yet conventional distance-based, mobility-based, and trust-based techniques adapt poorly to dynamic conditions, rely on subjective or fixed cluster head selection weights, and model the network as a simple pairwise graph that cannot capture the higher-order interactions of dense urban traffic. Reliable communication is further threatened by intrusions such as spoofing, packet dropping, and denial-of service attacks that cryptographic mechanisms alone cannot detect. This thesis addresses these coupled challenges by developing an intelligent, stability-aware, and secure clustering framework built on hypergraph-based spectral clustering, objective multi-criteria cluster head selection, and graph-based deep learning intrusion detection. The research pursues four objectives: to analyse existing VANET clustering algorithms, to develop scalable spectral clustering algorithms, to implement an efficient stability-based model benchmarked against the state of the art, and to analyse the reliability of the selected cluster heads. Guided by a bibliometric study of the field and an investigation of V2V and V2I communication reliability, two spectral clustering frameworks are proposed. The first represents the vehicular network as a weighted hypergraph, fixes the optimal number of clusters through an eigen-gap heuristic, forms clusters using an Improved Tensor Trace Maximization (iTTM) procedure over a modularity-based Laplacian, and elects cluster heads through an objective weighted Criteria Importance Through Intercriteria Correlation (CRITIC) decision over four criteria relative speed, connectivity level, link lifetime, and v consensus trust achieving 84% cluster head stability with only three clusters, together with higher throughput and reduced packet delay. The second framework retains the hypergraph foundation but introduces a modified Tensor Trace Maximization (mTTM) pipeline with Gaussian Mixture Model assignment and an Adaptive Neighbouring Behaviour-Incentivization (ANB-I) mechanism that rewards steady vehicles and penalizes erratic ones, achieving 75% stability with three clusters a 4.17% improvement over the hypergraph baseline and consistently low packet delay across sparse and dense networks. Both frameworks were evaluated on a realistic urban scenario derived from Connaught Place, Delhi, using SUMO generated mobility streamed into MATLAB. To secure this stable clustering, a Knowledge Graph-based Deep Convolutional Neural Network (KG-DeepCNN) Intrusion Detection System (IDS) is deployed at the roadside unit for intrusion-aware access control. The model combines knowledge graph and graph convolutional network embeddings with a weighted ensemble of four heterogeneous CNN tunnels, and was validated on the BoT-IoT, ToN-IoT, UNSW NB15, and NSL-KDD benchmarks, attaining accuracies of approximately 99%, 98%, 98%, and 99% respectively, with a 2.01% accuracy gain on UNSW-NB15 over comparable methods and composite cluster head reliability between 98.43% and 99.38%. Taken together, the results demonstrate that objective, behaviour-aware clustering on a scalable spectral foundation, protected by knowledge-driven intrusion detection, yields VANET communication that is simultaneously stable, reliable, and secure. By sustaining stable cluster heads, minimizing packet delay, and screening malicious access requests, the proposed frameworks support the timely delivery of safety critical messages, reduce signalling overhead, and contribute to safer and more efficient intelligent transportation systems.

2026-07-01T00:00:00Z SRIVASTAVA, SHIVAM Dahiya, Sonika (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23388 2026-10-16T06:34:44Z 2026-09-01T00:00:00Z

Title: COMPARATIVE ANALYSIS OF CNN VARIANTS AND VISION TRANSFORMERS FOR EARLY LUNG CANCER DETECTION Authors: SRIVASTAVA, SHIVAM; Dahiya, Sonika (SUPERVISOR) Abstract: Lung cancer continues to be a leading cause of death from cancer, and the prognosis largely relies on prompt and precise detection. Traditional methods of disease diagnosis require the interpretation of radiographs and histopathological images that require a lot of effort and suffer from observer variability. It has become crucial, therefore, to develop deep learning models in the field of computer aided diagnosis due to the ability of modern image models to extract discriminative features of interest and make decisions quickly. This thesis aims at evaluating several deep learning architectures and comparing their diagnostic capabilities for detecting early signs of lung cancer. One major goal of the thesis is to investigate the effect of deep learning model characteristics on diagnostic performance. These characteristics include model depth, parameter efficiency, feature reuse, transfer learning capability, and attention mechanisms. Architectures of CNNs and transformers will be reviewed with respect to the abovementioned characteristics, specifically Alex-Net, VGG16, ResNet50, DenseNet121, Inception Net, EfficientNetB0, Vision Transformer (ViT), and DeiT. An explanation and justification will follow for the choice of Vision Transformer over other model types. Preprocessing steps in the experimentation framework include resizing, contrast adjustment, normalization, augmentation, and splitting of the public datasets containing histopathology lung cancer The effectiveness of the algorithms is gauged using performance metrics like accuracy precision, recall, F1 score, the area under the receiver operating characteristic curve, and also confusion matrices. The results from the comparison suggest that deep networks, especially Inception-Net DenseNet121 Efficient-Net and transformer based networks, are promising approaches for lung cancer detection because of their discriminative power and generalization ability. There are other elements that get attention too, such as limitations like dataset biases, small clinical trial groups, computational costs, model interpretability, and the need for external verification before any clinical use, full stop. Overall, this thesis kind serves as an extension of the paper that…

2026-09-01T00:00:00Z RANJAN, HIMANSHU Rahul (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23377 2026-10-16T06:33:21Z 2026-05-01T00:00:00Z

Title: DIABETIC RETINOPATHY DETECTION USING DEEP LEARNING Authors: RANJAN, HIMANSHU; Rahul (SUPERVISOR) Abstract: Diabetic Retinopathy (DR) is a very critical eye disease that is a cause of blindness and vision problems globally and is one of the most serious diabetes-related eye diseases. It results from the damage to blood vessels in the retina from diabetes. The early diabetic retinopathy may not have obvious symptoms and diagnosis is very critical to prevent permanent loss of vision. A conventional diagnosis of DR is currently based on manually analysis of retinal images, conducted by an ophthalmologist, but this process is expensive, time taking and often impractical, especially in rural and resource-constrained healthcare settings where healthcare specialists are not readily available. Hence, the need of intelligent and automated diagnostic system for early detection and accessibility to healthcare system is of utmost necessity.This research introduces theEye-RDNN: a deep learning framework designed for the classification of diabetic retinopathy in fundus imagesof retina with several classes. The suggested model features convolutional neural network, along with transfer learning, to automatically classify various stages of DR in retinal images. During the training process, image resizing, normalization, sample augmentation (rotation, zooming, brightness enhancement, and flipping), and other techniques were applied to the image set to improve the performance of the model. These procedures have enhanced the diversity of datasets, reduced overfitting, and boosted the feature extraction capacity of the models.The entire model was developed on TensorFlow/Keras models using Google Colab with help of support of GPUs for efficient computing and faster trainings. The choice of the base architecture was EfficientNet-B0 due to its efficient computations and excellent feature extraction property which proved to be adequate for the purpose of medical image classification. The accuracy of the trained model in classification of multiple stages of DR was highly satisfactory for images of the healthy retina as well as 6 severitylevels of diabetic retinopathy.The system designed can help the ophthalmologist carrying out the retinal screening faster and efficient, which will reduce the manual workload and the risk of delayed diagnosis. This sort of automated systems are especially helpful in far-off health care areas the place skilled ophthalmologists are scarce. Furthermore, the proposed approach could potentially be applied in large scale DR screening initiatives and enhance the detection of early DR. Additional features for the model could be employed to get even the better results, such as increase in the count of patients included in the data set analyzed and including diversity and varieties in the data, using more sophisticated and complex deep learning algorithms and booking the model for clinical validation, which could lead us to a higher rate of correct and accurate diagnosis and an increase of clinical practices that could be applied from the model.

2026-05-01T00:00:00Z