There is an increased adoption of electronic health record systems by a variety of hospitals and medical centers. This provides an opportunity to leverage automated computer systems in assisting healthcare workers. One of the least utilized but rich source of patient information is the unstructured clinical text. In this work, we develop CATAN, a chart-aware temporal attention network for learning patient representations from clinical notes. We introduce a novel representation where each note is considered a single unit, like a sentence, and composed of attention-weighted words. The notes in turn are aggregated into a patient representation using a second weighting unit, note attention. Unlike standard attention computations which focus only on the content of the note, we incorporate the chart-time for each note as a constraint for attention calculation. This allows our model to focus on notes closer to the prediction time. Using the MIMIC-III dataset, we empirically show that our patient representation and attention calculation achieves the best performance in comparison with various state-of-the-art baselines for one-year mortality prediction and 30-day hospital readmission. Moreover, the attention weights can be used to offer transparency into our model's predictions.
Deep transfer learning is a popular choice for classifying monochromatic medical images using models that are pretrained by natural images with color channels. This choice may introduce unnecessarily redundant model complexity that can limit explanations of such model behavior and outcomes in the context of medical imaging. To investigate this hypothesis, we develop a configurable deep convolutional neural network (CNN) to classify four macular disease conditions using retinal optical coherence tomography (OCT) images. Our proposed non-transfer deep CNN model (acc: 97.9%) outperforms existing transfer learning models such as ResNet-50 (acc: 89.0%), ResNet-101 (acc: 96.7%), VGG-19 (acc: 93.3%), Inception-V3 (acc: 95.8%) in the same retinal OCT image classification task. We perform post-hoc analysis of the trained model and model extracted image features, which reveals that only eight out of 256 filter kernels are active at our final convolutional layer. The convolutional responses of these selective eight filters yield image features that efficiently separate four macular disease classes even when projected onto two-dimensional principal component space. Our findings suggest that many deep learning parameters and their computations are redundant and expensive for retinal OCT image classification, which are expected to be more intense when using transfer learning. Additionally, we provide clinical interpretations of our misclassified test images identifying manifest artifacts, shadowing of useful texture, false texture representing fluids, and other confounding factors. These clinical explanations along with model optimization via kernel selection can improve the classification accuracy, computational costs, and explainability of model outcomes.
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：