Abstract:
Kidney stone disease is a prevalent urologic problem, and recurrence rates are high, thus it
impacts patients' quality of life and health care costs. Diagnosis in most methods is imaging and invasive procedures; hence, they are costly and time-consuming. This study aims at meeting this critical need for a low-cost and non invasively prediction model that would be used in the detection of the presence of kidney stones and assessment of the risk of recurrence based
on urinalysis variables such as specific gravity, pH, osmolarity, and calcium levels.
To achieve this, we have developed an integrated deep learning-based predictive model with
techniques of explainable AI, which will enhance transparency. A sequential neural network
design was trained using techniques such as data normalization and class weighting on
urinalysis feature data to handle the imbalance in the dataset. We incorporated explainable AI visualization tools, such as SHAP or LIME.
Initial results showed an accuracy of around 50.06%, it would be clear where the work needed to be improved in enhancing the model performance classification between stone versus no stone. Further refinement in hyperparameters and feature engineering is expected to increase this accuracy, yet at the same time, explain the model better. This will yield advances in non-invasive renal pathology diagnosis, securing earlier and easier assessment of kidney stone risk.
