Machine Learning for the Real World: Applications and Insights

A Prediction System for Early Identification of Students at Risk of Mental Health Issues

2026-02-20T09:33:17+00:00

Mental health among university students in India is a critical public health concern, with studies indicating that over 37% of students exhibit symptoms of moderate to severe distress, profoundly impacting academic success and leading to high dropout rates. This chapter details the design, implementation, and evaluation of the Mental Health Risk Prediction System (MHRPS), a proactive, data-driven software solution developed to address this challenge. The MHRPS leverages machine learning to analyse a holistic set of student data—encompassing academic performance, behavioural engagement, and self-reported psychological metrics from instruments like the PHQ-9 and GAD-7. The system's data pipeline, built entirely in R (version 4.5.1), uses the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance, a common problem in health data. It then employs a fine-tuned random forest model to classify students into 'Low', 'Medium', or 'High' risk categories. The resulting model achieved 99.87% accuracy and an Area Under the Curve (AUC) approaching 1.0, demonstrating a robust ability to distinguish between risk levels. Though this performance is contextualised within the study's methodology. Analysis revealed that direct self-reporting on items like self-harm thoughts and standardised questionnaire scores were the most powerful predictors. Actionable insights are presented via an interactive R Shiny dashboard, providing counsellors with individualised risk profiles. The entire workflow, from daily data synchronisation with Google Sheets to monthly model retraining, is fully automated using the Windows Task Scheduler. Deployed on a standard Windows PC using an entirely open-source software stack, the MHRPS is designed as an economically and technically feasible solution for Indian universities, offering a powerful tool to foster a more supportive and mentally healthy academic environment.

Critical Care Analytics: Insights from the Emergency Room

2026-02-20T09:36:02+00:00

Emergency departments (EDs) face unpredictable patient surges, fragmented data systems, and operational inefficiencies that delay clinical decision-making and increase patient wait times, staff burnout, and risk of adverse outcomes. This chapter aims to introduce the Critical Care Data Analysis (CCDA) framework—an integrated machine-learning and time-series forecasting system with real-time system-level monitoring—to optimise emergency room (ER) operations. CCDA combines electronic health records, staffing schedules, and environmental data to predict patient volumes, forecast resource shortages, and detect staff burnout. Initial simulations demonstrated measurable benefits, including a 20 % reduction in average wait times, a 15 % improvement in staff allocation efficiency, and a 10 % decrease in equipment-shortage incidents. A real-time Power BI dashboard provides actionable alerts and performance diagnostics. This scalable, data-centric approach offers a robust decision-support tool for improving ER efficiency, resilience, and clinical outcomes.

A Stable and Adaptable Machine Learning Framework for Phishing Detection

2026-02-20T09:40:05+00:00

Background: Phishing contributes to over one-third of security incidents globally, highlighting the urgent need for robust detection systems.

Aims: The aim of this study is to design and validate a phishing detection system that ensures accuracy, adaptability, and real-time deployment suitability. The system targets institutional and enterprise-level use, focusing on overcoming the shortcomings of traditional rule-based and blacklist approaches.

Study Design: An experimental research study was conducted to evaluate multiple machine learning algorithms for phishing detection. The study adopted a comparative design to identify the most stable and efficient model.

Place and Duration of Study: The work was carried out at the Department of MCA, Acharya Institute of Technology, Bangalore, India, between July 2025 and September 2025.

Methodology: A dataset of SMS messages, consisting of 5,559 messages labelled (phishing and legitimate), was pre-processed using tokenisation, stop-word removal, and vectorisation (TF-IDF and BoW). Lexical, structural, statistical, and semantic features were engineered. Six classifiers—Multinomial Naïve Bayes (MNB), Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbours (kNN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)—were trained and evaluated using Accuracy, Precision, Recall, and F1-score. Cross-validation was applied for stability testing. A Django-based web interface was implemented for real-time predictions.

Results: The proposed method uses many algorithms of machine learning with feature engineering to find phishing sites. Support Vector Machine achieved the best stability with 99.99% Accuracy, 98.99% Precision, 99.12% Recall, and 99.05% F1-score. MNB, kNN, and LSTM achieved near-perfect results, while CNN performed relatively lower (Accuracy 91.02%). Real-time system testing showed an average response time of 0.05 seconds per message.

Conclusion: The proposed phishing detection system demonstrates strong accuracy, efficiency, and adaptability. Its lightweight design and real-time performance make it suitable for deployment in institutional servers, email systems, and organisational networks, providing an effective defence against evolving phishing attacks.

Diabetic Retinopathy Detection Using CNN (Convolutional Neural Networks)

2026-02-20T09:43:32+00:00

Diabetic retinopathy (DR) is a leading preventable cause of blindness worldwide.

It often has advanced progression prior to exhibiting significant symptoms. Early detection and accurate grading of DR are important for treatment and also for the prevention of visual loss. The severity levels of No DR, Mild, Moderate, Severe, and Proliferative DR have been categorised for retinal fundus images for this analysis using a Convolutional Neural Network (CNN) model. This chapter investigates the application of CNN model to detect Diabetic retinopathy. Grad-CAM or gradient-weighted Class Activation Mapping will be applied thereafter to improve interpretability by highlighting clinically relevant features like haemorrhages, microaneurysms and exudates. Using preprocessing techniques such as scaling, normalisation, and augmentation improved the model's ability to generalise on both the EyePACS and Messidor datasets. The CNN produced a macro-averaged recall of 92%, accuracy of 93%, precision of 91%, and ROC-AUC of 0.95. In other words, dividing the remaining probabilities by No DR (23.9%), Mild (7.5%), Severe (6.5%) and Moderate (3.8%) respectively, while there was a single case of Proliferative DR that had a 58.3% confidence. To deploy it in real time, a simple web application with Flask was developed to get Grad-CAM overlays and predictions in seconds. The system is a viable option for early DR screening due to its accuracy, interpretability, and usability, particularly in low-resource settings and clinical settings.

Previous studies have initially demonstrated the effectiveness of deep learning-based systems for the detection and grading of diabetic retinopathy [1,6,11,15], which further encourages the adoption of CNN techniques in clinical screenings.

EVision India: A Smart Dashboard for Electric Vehicle Sales and Buyer Decisions

2026-02-20T09:46:44+00:00

In India, the increasing adoption of electric vehicles (EVs) due to environmental, energy security, and government policy factors has raised numerous research possibilities in terms of sales trajectories, infrastructure development, consumer acceptance, and sales forecasting. However, much of the existing literature remains siloed: independent studies approach the subject singularly (e.g., some examine only EV adoption barriers while others examine forecasting). This review looks to synthesise recent research in which machine learning models, time-series forecasting, new federated learning applications, and policy frameworks converge. Research gaps will be polled or inferred through an analysis of the current literature, such as limited visualisation tools developed by authors with a history in industry, a lack of comparative analysis of state-based EV infrastructure planning, and weak consumer-facing decision support. We summarise the directions offered by an example, EVision India, a smart dashboard that combines sales analytics, infrastructure layer mapping, forecasting models, and a recommendation engine to provide key stakeholders, including policymakers, manufacturers, and consumers, a roadmap to navigate EV purchases, inform government policies, and maximise EV supply chain efficiencies. The paper/authority ends with suggestions for future areas of EV research, including IoT-based, real-time data, explainable artificial intelligence (AI) to improve consumer trust, and platform scalability for use in EVs in global markets.

Face Recognition for Missing Person Identification

2026-02-20T09:49:31+00:00

Every year, thousands of people, including both children and adults, go missing for many different reasons, including behavioural and mental health, natural disasters, accidental separations, and human trafficking. Facial recognition has already demonstrated its usefulness across a variety of fields in both humanitarian relief and security and surveillance. The aim of this project was to create an automated system that would orchestrate the identification of missing persons, utilising face recognition with real-time recognition and detection, eliminating human error and establishing minimal time for a response. This study provides an applied system that uses computer vision, machine learning, and database management in law enforcement agencies, non-profits, and families. The work was completed at the Department of MCA, Acharya Institute of Technology, Bangalore, India, from July 2025 to September 2025, executing experiments with benchmark datasets, test datasets, and live camera feeds. For face detection, a Haar Cascade Classifier was applied, while Local Binary Pattern Histogram (LBPH) was utilised for recognition. In addition, compared experiments were implemented with YOLO (version 5) for real-time detection and HOG + SVM for feature-based recognition. Pre-processing (resizing, normalisation, and histogram equalisation) was done before feature extraction, and the images were stored in MongoDB for comparison against missing persons cases. Once a match was established for a missing person found, an established Email API facilitated automated alerts to authorities and registered family members. The conventional approach (Haar + LBPH) reached identification accuracy from 82–88% in good light and frontal angle, YOLO (v5) surpassed, showing slightly over 98%, while providing balanced precision/recall, HOG + SVM had similar results around an identical 96%. The system was capable of real-time performance on a CPU machine, making it great for low-cost and light deployment.

The proposed system shows automated alerts combining anthropometric (Haar, LBPH and HOG + SVM) methods and advanced deep learning (YOLO, CNN-based architectures (ResNet, ArcFace and FaceNet) to provide a reliable and scalable solution for missing persons identification. The system reduces human effort, increases response time velocity and opens up a platform for development in the area of deep learning, multimodal data fusion, and mobile-based applications.

Adaptive Study Planning Using Machine Learning for Personalised Learning

2026-02-20T09:52:47+00:00

Personalised learning, which refers to an educational paradigm that allows individual students to dictate content and schedule of their learning, has begun to assume dominance as a new level of education transformation, and has been shown to promote better retention (by about 20%) when plans reflect the learner's states. This study delineates the implementation of a machine learning-based approach toward personalised study plan generation and personalised scheduling that takes into account individual lifestyle factors and attention profiles. This experimental study was carried out in the Department of MCA, Acharya Institute of Technology, Bengaluru, India (January 2024 - August 2024) and applied Random Forest regression (RF) and KMeans cluster analysis (k-means) to a dataset of student-specific schedules based upon eight lifestyle variables: sleep/hours, screen/hours, stress level, mental health rating, frequency of exercise, motivation level, time management score and daily study hours. The experimental study underwent a preprocessing stage; Random Forest exhibited strong performance (MAE of 3.46 and R-squared of 0.97) and identified importance ranking predictors of attention; participants' sleep and stress have been identified as the highest ranked predictors of attention in Random Forest. K-Means managed to classify learners into Low, Medium, and High focus profiles (k=3 and a silhouette score of 0.53 Which is moderately acceptable), which led to scheduled Pomodoro sessions being allocated on study plans. The Pomodoro sessions were allocated during study hours as the Largest Remainder Method. The system development utilised Flask and SQLite and was able to create a study plan in less than two seconds; system adaptations positively impacted feedback, enabling learners to complete 15% more of their study plans. The results presented here indicate that combining variables related to lifestyle will permit more personalised and effective study plans versus traditional methods. The main objective of the system presented was to have a scalable, responsive, and privacy-respecting solution to increase student focus, decrease procrastination behaviours, and improve learning outcomes, while also providing institutions with systems that can assist in adaptive, data-informed educational practices in the future.