Adaptive neural tree exploiting expert nodes to classify high-dimensional data

Achalasia phenotypes and prediction of peroral endoscopic myotomy outcomes using machine learning

OBJECTIVES

High-resolution manometry (HRM) and esophagography are used for achalasia diagnosis; however, achalasia phenotypes combining esophageal motility and morphology are unknown. Moreover, predicting treatment outcomes of peroral endoscopic myotomy (POEM) in treatment-naïve patients remains an unmet need.

METHODS

In this multicenter cohort study, we included 1824 treatment-naïve patients diagnosed with achalasia. In total, 1778 patients underwent POEM. Clustering by machine learning was conducted to identify achalasia phenotypes using patients' demographic data, including age, sex, disease duration, body mass index, and HRM/esophagography findings. Machine learning models were developed to predict persistent symptoms (Eckardt score ≥3) and reflux esophagitis (RE) (Los Angeles grades A-D) after POEM.

RESULTS

Machine learning identified three achalasia phenotypes: phenotype 1, type I achalasia with a dilated esophagus (n = 676; 37.0%); phenotype 2, type II achalasia with a dilated esophagus (n = 203; 11.1%); and phenotype 3, late-onset type I-III achalasia with a nondilated esophagus (n = 619, 33.9%). Types I and II achalasia in phenotypes 1 and 2 exhibited different clinical characteristics from those in phenotype 3, implying different pathophysiologies within the same HRM diagnosis. A predictive model for persistent symptoms exhibited an area under the curve of 0.70. Pre-POEM Eckardt score ≥6 was the greatest contributing factor for persistent symptoms. The area under the curve for post-POEM RE was 0.61.

CONCLUSION

Achalasia phenotypes combining esophageal motility and morphology indicated multiple disease pathophysiologies. Machine learning helped develop an optimal risk stratification model for persistent symptoms with novel insights into treatment resistance factors.

Theory of adaptive SVD regularization for deep neural networks

Deep networks can learn complex problems, however, they suffer from overfitting. To solve this problem, regularization methods have been proposed that are not adaptable to the dynamic changes in the training process. With a different approach, this paper presents a regularization method based on the Singular Value Decomposition (SVD) that adjusts the learning model adaptively. To this end, the overfitting can be evaluated by condition numbers of the synaptic matrices. When the overfitting is high, the matrices are substituted with their SVD approximations. Some theoretical results are derived to show the performance of this regularization method. It is proved that SVD approximation cannot solve overfitting after several iterations. Thus, a new Tikhonov term is added to the loss function to converge the synaptic weights to the SVD approximation of the best-found results. Following this approach, an Adaptive SVD Regularization (ASR) is proposed to adjust the learning model with respect to the dynamic training characteristics. ASR results are visualized to show how ASR overcomes overfitting. The different configurations of Convolutional Neural Networks (CNN) are implemented with different augmentation schemes to compare ASR with state-of-the-art regularization methods. The results show that on MNIST, F-MNIST, SVHN, CIFAR-10 and CIFAR-100, the accuracies of ASR are 99.4%, 95.7%, 97.1%, 93.2% and 55.6%, respectively. Although ASR improves the overfitting and validation loss, its elapsed time is not significantly greater than the learning without regularization.