Classification of Patient Satisfaction among People Living with HIV (PLHIV) Accessing Healthcare Services in Zambia: A Machine Learning-Driven Approach

This study focused on how Zambian HIV treatment facilities employed machine learning methods, specifically sentiment classification techniques, to systematically sort patient happiness based on qualitative healthcare data. Manual analysis of patient feedback takes a lot of time, money, and effort and frequently produces biased results. More importantly, this reliance on outdated methods makes it harder for us to understand and fix major problems in the healthcare system, such as persistent stigmatisation and unpredictable drug shortages. This study's main objective was to develop and compare machine learning models, including deep learning and traditional classifiers, to accurately and objectively classify patient satisfaction sentiment, thereby addressing the limitations of manual qualitative analysis in Zambian HIV healthcare. From January to November 2023, a full survey was done at thirty healthcare facilities in nine administrative districts in Zambia's Southern Province. The study involved gathering both qualitative and quantitative data from 3,052 People Living with HIV (PLHIV) who received care. The methodological framework included extensive text preprocessing steps, polarity quantification using TextBlob, and the use of different machine learning architectures, such as logistic regression, Support Vector Machines, Naive Bayes, Random Forest, Gradient Boosting, Long Short-Term Memory networks (LSTM), Bidirectional Encoder Representations from Transformers (BERT) fine-tuning, and Convolutional Neural Networks. Examination of patient feedback showed that 87% of the answers were good, 5% were neutral, and 8% were negative. Thematic evaluation highlighted positive aspects like polite staff and efficient service but also revealed serious problems, such as long wait times, a lack of medication, and staff behaviour difficulties. The fine-tuned BERT model did very well on classification tasks, getting 97% accuracy and matching weighted precision, recall, and F1-scores. The CNN model, which used word embeddings, did just as well on these same benchmarks, with an accuracy rate of 96%. This was a big improvement over traditional machine learning methods, which had accuracy rates of 82% to 91%. This study shows that adding machine learning-based sentiment analysis to normal electronic health record systems at facilities has a lot of potential. It would create a scalable, objective, and responsive framework for improving the quality of HIV care on an ongoing basis. Keywords: Zambia, HIV care, patient feedback, public health, machine learning, sentiment analysis, natural language processing, deep learning, model evaluation, healthcare quality