Anchal Garg

In healthcare communication, technical jargon in medical reports is hard for patients to understand. This lack of clarity may lead to follow-up appointments and non-adherence to treatment, with adverse effects on patient outcomes. Despite its profound impact, there are few practical solutions to bridging this communication gap. To address this, we developed an AI-powered radiology report analyzer. Using Natural Language Processing, our tool identifies complex medical jargon in uploaded reports and provides simplified definitions along with relevant visual aids. These terms are clickable, also allowing patients to query using a chat-based interface, transforming static reports into dynamic and user-friendly experiences. We tested our tool on radiology reports from The NationalRad Sample Reports on readability, understandability, and actionability. Compared with the standard reports, which were at a university reading level on average and graded "difficult," our translated versions were at a seventh-grade reading level. Our tool scored better on readability and understandability metrics: Flesch‐Kincaid Grade Level (7.9 ± 1.2 vs 12.1 ± 0.3), Gunning‐Fog Score (9.2 ± 1.1 vs 16.4 ± 0.2), and Coleman‐Liau Index (9.9 ± 0.9 vs 15.6 ± 1.1), all statistically significant (P < .05). The tool also indicated higher PEMAT-P scores in understandability (93.8% vs 56.4%) and actionability (45.0% vs 30.6%). The pilot study demonstrates the effectiveness of our system in augmenting health literacy and enabling informed decision-making and thus promoting patient-centered care. Future development will focus on expanding visual aid coverage for non-anatomic terms for greater comprehension.

Understanding and tracking emotions is vital for improving athlete recovery, performance, and mental wellbeing. Yet, most existing systems lack real time analysis and use overly complex emotion classifications that are hard to apply in practice. This paper presents an emotion analysis system that uses advanced techniques in Artificial Intelligence and Natural Language Processing. A fine-tuned deBERTa-v3 model, trained on the GoEmotions dataset and mapped to seven Ekman emotion categories is used for simplified emotion classification. The users' complete surveys before and after running and the system provides real-time emotion labels with confidence scores. These results are used to generate a dashboard of weekly emotion trends. On a validation split of the official GoEmotions dataset, the model achieved 69.73% accuracy and 61.79% F1 score. The paper also compares this model with baseline methods. The system employs a web-based interface and an API-driven backend to facilitate real-time evaluation of the emotion classification model.

The boom in the real estate market necessitate easier and modern prediction models to accurately estimate properties' values. In this research, seven machine learning algorithms for house price prediction on residential houses in Bolton, Greater Manchester, were evaluated to cover the research gap in local UK market research. Following the CRISP-DM guideline, analysis of 13,933 transactions of 2,930 unique properties over a period of 12 months based on 17 property attributes with GDPR compliance were made. The pipeline for preprocessing end-toend comprised correlation analysis, detection of outliers using IQR, log transformation for skewness correction, and Min-Max normalization. Seven different models were experimented with: Linear SVR, Polynomial SVR, RBF SVR, K-Nearest Neighbors, Decision Tree, Random Forest implementations (RF-100, RF-200, RF-500), and XGBoost. Performance assessment with 70/30 traintest split showed Random Forest with 500 estimators yielded superior results ( R^{2}=0.93 , MAE =\mathbf{0. 1 0 8 6} , MSE =\mathbf{0. 0 3 3 2} , RMSE =0.1822 ), then XGBoost ( R^{2}=0.92 ). Total Living Area was the superior predictor ( r=0.67 ), confirming rudimentary real estate appraisal fundamentals. These findings offer proof of the capacity of machine learning to enhance property valuation accuracy, with implications for applied utilization by real estate professionals, financial institutions, investors, and policymakers in regional market analysis.

Vaccination is a key to public health and safety, but adverse vaccine side effects threaten its safety and monitoring. Traditional pharmacovigilance systems are characterised by flaws such as underreporting, delays, and data fragmentation. This paper presents a holistic review of the role of artificial intelligence (AI) in vaccine side effect detection and prediction. Methodologies like machine learning, deep learning, and natural language processing are examined, with a close look at their application to real-world data sources like electronic health records, social media, and vaccine registries. Evidence from the COVID-19, HPV, and influenza vaccines shows that AIs can detect vaccine side effects, with each of the AI models achieving a high accuracy of (AUC 0.91, F1-score 0.903) in vaccine side effect detection. However, despite the promising benefits, some challenges such as data quality, model generalisation, and ethical challenges still exist. This paper also suggests that future research directions should include vaccinology, federated learning, and explainable AI for real-time and personalised risk assessment. Conclusively, this paper brings to the fore the numerous benefits that AI offers in pharmacovigilance, showcasing its transformative potential for proactive vaccine safety surveillance.

Vehicle cloning, the act of duplicating licence plates and in some cases whole car is an emerging threat which presents a challenge that has not been catered for in existing solutions. Although the Automatic Licence plate Recognition (ALPR) system has been at a mature stage for a while, this unique problem is not addressed, and we have not come across any commercial solution either. This paper is first to address two key areas in enhancement of ALPR system. First the evaluation metric was developed to address unfairness of accuracy of string level match enhancing accuracy from 43% to 96%. Secondly, we introduced a novel framework using 'Usual Route Identification' algorithm to analyse a Vehicle's usual journey which helps to flag if unusual routes are taken or located outside off a set radius indicating a potential case of vehicle cloning. This study will show how appropriate metric can enhance the capability of ALPR systems, proposes a functionality that aids law enforcement authorities, and adds to academic knowledgebase.

Laughter, particularly self-induced laughter, has known therapeutic benefits. However, its acoustic-emotional dynamics remain underexplored compared to spontaneous or social laughter. This study presents a novel approach to classifying post-laughter emotional states (positive vs. neutral) based on both global and time-segmented MFCC features—including delta and delta-delta coefficients—extracted from recordings of 126 participants under controlled conditions. Emotion labels were obtained via immediate self-reports to minimize subjective bias. Analysis revealed that acoustic features from the later segments of laughter sessions are most predictive of emotional outcome. Among several models evaluated, BiLSTM achieved the highest performance (86.67% accuracy, F1 score = 0.87, and AUC = 0.96), indicating its strength in modeling temporal patterns in laughter. These findings not only advance emotion recognition from nonverbal cues but also offer insights for designing AI systems capable of generating or interpreting context-sensitive, emotionally relevant laughter—such as in therapeutic or assistive human-computer interactions.

Laughter is known to improve mental well-being. Laughter is generally categorized into self-induced and externally induced laughter, and there is a lack of empirical evidence differentiating the two. There are limited studies on the use of brain signals to differentiate frequency patterns related to these two types of laughter, and to explore their role in mental health and well-being. This study aims to address this gap using brain frequency wave responses. Brain frequency data were collected from fifty participants using a Muse headband. MNE-Python, Independent Component Analysis, and time-frequency were used for exploratory analysis. Machine learning and deep learning techniques (Random Forest, Gradient Boosting, LSTM, and Logistic Regression) were used to classify EEG trends. Random Forest revealed greater accuracy of 74%. Brainwave trends differed notably between the two types of laughter. Brain signals during Self-induced displayed prominent beta and gamma responses, while externally induced showed significant alpha and theta values. Thus, the self-induced laughter has a stronger impact on brainwaves connected to cognitive engagement and mental health compared to externally induced laughter. The research provides evidence that laughter can be prescribed to improve mental health and well-being. This research aids the utilization of EEG data for laughter analysis and unlocks paths for future studies into the therapeutic use of laughter for mental health advancements.

Medical reports can contain technical vocabulary, abbreviations, and technical terms that are barriers to patient comprehension, with impacts on health literacy and decision-making. This article describes the design and evaluation of an AI system to simplify automated medical reports, generate visual aids, and interactive question answering. The system includes OCR for text extraction, NLP pipelines for terminology simplification, and a Q&A module using large language models (LLMs). Three models, ChatGPT, ClinicalBERT, and DeepSeek, were comparatively evaluated on four tasks: term extraction, explanation quality, term-to-image mapping, and relevance in dialogue. DeepSeek performed the best among all models with 0.92 F1-score, 0.84 BLEU, and 88% visual mapping success. A hybrid pipeline integrating BioBERT with generative LLMs improved accuracy by 12% over single-model baselines. Findings indicate that a blend of domain-specific extractors and generative models provides a strong methodology for enhancing patient-focused medical communication.

This research examines the customer churn issues in the telecommunications industry, which significantly impact companies by leading to revenue loss and high costs associated with acquiring new customers. Therefore, effective prediction and preventing churn are still significant challenges in the industry. This research utilises deep learning techniques to implement a model that can predict customer churn. The dataset used is the "Customer Churn Prediction" dataset from Kaggle, which contains 100,000 customer records with 100 features, including demographic information, service usage patterns, billing details, and customer behaviour metrics collected over a 12-month period from a major telecommunications provider. Moreover, the Genetic Algorithm is used as a feature selection technique to find the best features from the dataset and help to improve the model's performance. The most crucial factors that impact churn, such as customer usage patterns, network quality, and engagement with customer care, were identified. Moreover, the implemented Deep Learning model, Temporal Convolutional Network, used train and test data to calculate several metrics. Hyperparameters for TCN are used to improve model performance. The results are an accuracy of 70.4%, a sensitivity of 71%, and an AUC of 0.8. These results outperform the single models, such as Extreme Gradient Boosting (XGBoost), Random Forests (RF), Light Gradient Boosting Machine (LGBM), Logistic Regression (LR), K-nearest neighbours (KNN), Support Vector Machines (SVM), Naïve Bayes, and AdaBoost. The suggested model helps telecom companies keep their customers and reduce churn rates by providing a practical solution.

This research applies natural language processing, machine learning, and deep learning techniques for the real-time detection of Domestic Violence using indicators from X (Twitter) data. For a dataset of over 2.9 million tweets, the study proposes a multi-step computational pipeline comprising data gathering, preprocessing, feature engineering, and model evaluation. The study also addresses data imbalance using SMOTE and uses topic modeling and TF-IDF feature extraction for meaningful representation. The methodology integrates classic models such as Decision Trees and Naïve Bayes with contemporary architectures such as LSTM, CNN, and a hybrid CNN-LSTM model. Comparative analysis using accuracy, precision, recall, and F1-score shows LSTM to be superior to all models. The findings point to the efficacy of deep learning for detecting subtle indicators of domestic violence, and the moral responsibility that accompanies automated detection. The research offers a baseline model for leveraging AI to inform early intervention and policy development, for the creation of safer online environments, and an evidence-based community response.