Would You Trust an AI Doctor? Building Reliable Medical Predictions with Kernel Dropout Uncertainty