Development and Evaluation of Deep Learning Models for Healthcare Classification
This individual technical assessment focuses on the design, development, analysis and evaluation of a deep learning solution for a healthcare-related classification problem. Students select one of two provided scenarios: Polycystic Ovary Syndrome (PCOS) detection using ultrasound images or heartbeat classification using electrocardiogram (ECG) signals. The objective is to develop an appropriate deep learning approach and demonstrate critical understanding of the complete machine learning workflow, from initial data exploration through to model evaluation and reflection. Students may either design and train a deep learning model from scratch or customise and fine-tune an existing pre-trained architecture. The complete work is presented through a single Jupyter Notebook integrating Python code, technical discussion, results and visualisations. The notebook must clearly define the selected healthcare problem, explain its significance, justify methodological and architectural choices, and critically evaluate the resulting solution. The first stage involves exploratory data analysis and preprocessing, including investigation of class distributions, data imbalance and relevant patterns. Students prepare the data through techniques such as normalisation, augmentation, train-validation-test splitting and appropriate handling of class imbalance. This is followed by model design, training, validation and hyperparameter tuning, with the architecture selected according to the characteristics of the data and classification task. Model performance must then be evaluated using appropriate classification measures, including precision, recall, F1-score, ROC curves and area under the curve (AUC). The developed model should also be compared against suitable benchmark approaches, which may include traditional machine learning algorithms or alternative deep learning architectures. This comparison should identify the relative strengths and limitations of the proposed solution. The final component requires clear visual presentation and critical reflection on the complete modelling process, including limitations, challenges and opportunities for improvement. Importantly, grading prioritises methodological rigour, analytical depth and critical evaluation rather than simply achieving the highest predictive accuracy. Overview word count: approximately 330 words. AI restriction: this brief only permits automated AI tools for spelling and grammar checking. It explicitly prohibits tools such as ChatGPT, Gemini or Copilot from authoring assessment text or code; any permitted AI use must also be acknowledged.
Read Model Answer →