Automated Enteropathy: Discovering the Potential of Machine Learning in Environmental Enteropathy

Advances in computational techniques have driven countless recent developments in both research and clinical capacity. Supervised Machine learning[1], a term for computational models that learn to provide specific input/output from training data of that same form, has already been the focus of multiple efforts to advance clinical practice, in particular for use in the analysis of images such as in assessment of radiology and histology[2-4].

Sana Syed et al at the University of Virginia are leaders in the US investigation of environmental enteropathy (EE), otherwise known as environmental enteric dysfunction (EED), a disease felt to be the major global cause of stunted growth[5]. EE/EED is endemic to developing nations and felt to be linked to recurrent enteric infection in areas without access to clean water. With a subclinical presentation and no reliable biomarker[6], EE/EED requires a systematized approach to interpreting histologic evidence, an area of utmost importance to ongoing work in addressing a condition affecting hundreds of millions of children. Dr. Syed previously collaborated in several other parts of this effort, not least of which included serving on a team attempting to standardize histopathological markers of EE/EED across regions[7].

In this issue, Syed et al explore use of a more intricate model of machine learning known as deep learning, deploying tools called “Convolutional Neural Networks” (CNN) to assess digitized images of histological samples and ideally to differentiate between normal tissue and enteropathy caused by EE/EED and Celiac Disease (CD). Deep Learning is a simple, flexible, and surprisingly effective set of modeling components that has received significant investment in recent years in software, hardware, and academic research[1]. These algorithms are adept at image analysis, as they remove the need for “feature engineering” wherein researchers hand-design low level features used by the model to identify specific components of the image. The implications of success in automating initial histopathological screening are profound for clinical practice broadly, but specifically in the context of EE/EED in which trained personnel can often be a rate limiting step in the areas in which the disease is most prevalent. If successful, this technique and other methods can be applied more broadly to well characterized histopathological conditions, potentially leading to a screening tool for rapid characterization of an intestinal biopsy. Such a tool would have immense clinical impact, improving turnaround time and reliably identifying normal biopsy samples and referring abnormal samples for confirmation by an experienced pathologist. With regard to EE/EED, an automated scoring system provides a high utility tool for future research, allowing for rapid, cost-effective, and standardized histologic scoring to serve as correlation with novel efforts in assessing the impact of shifts in the microbiome or RNAseq efforts in affected patients[8].

Convolutional models such as the authors employ have demonstrated substantial success in image analysis, and this article describes an excellent first step in deploying it for histopathological analysis. Syed et al have achieved impressive results, evaluating three different models of deep learning (shallow CNN, resNET50 and multi-zoom resNET50) to generate a combined accuracy of 98.3% in an exclusive classification task between EE/EED, CD, and control samples. They also undertook an effort using the previously utilized “Grad-CAM” technique to alleviate the “black box” issue, the fact that while a system may be highly accurate, the reasoning behind its decision-making typically remains vague. The ability to characterize the decisions that lead to a classification is both reassuring and intriguing, as it may actually create the ability to identify new histological markers that are either subtle or challenging for pathologists to identify.

The generated accuracy of 98.3% is impressive, but comes with some caveats. In addition, when dealing with small datasets such as the one used in this work, it is common for non-generalizable factors or irrelevant idiosyncrasies to give deep learning models artificially elevated performance numbers[9]. In other words, with a small N it is possible for to create false pathways based on irrelevant findings but yielding correct answers. Due to limitations of available biopsy samples, the investigators were not able to provide a large dataset to their CNNs to learn EE characteristics. And while their accuracy is impressive, with a low N (only 48 available EE biopsies) it is hard to determine the significance and reproducibility of this finding with novel samples. This issue will be rapidly remedied as the investigators apply their model more broadly.

By expanding this work and continuing to provide training material from different sources, Syed et al will be able to estimate the accuracies of these methods to higher statistical significance. The use of combined model architectures in this study is also of note, as it exemplifies a technique that may help sustain the high accuracy of this approach as the N increases. Combining models falls under the term “mixture of experts,” a method to more reliably characterize images by using a principle similar to the “wisdom of the crowd.” By allowing multiple models to assess data, their findings can be aggregated to essentially create a meta-analysis of their conclusions. This approach has shown substantial capability[10], and training additional models to support an expanded “mixture of experts” may be of use as their dataset expands.

In summary, this is an important step in the process of developing automated screening tools capable of providing true decision support to our colleagues in pathology. The implications of developing this specific capability in EE/EED are profound, offering both an expansion of histopathological services in developing nations as well as possible research tools for improving our understanding of a condition impacting hundreds of millions of children