Improvement of adenoma detection rate by two computer-aided colonic polyp detection systems in high adenoma detectors: a randomized multicenter trial

BACKGROUND

This study aimed to evaluate the benefits of a self-developed computer-aided polyp detection system (SD-CADe) and a commercial system (CM-CADe) for high adenoma detectors compared with white-light endoscopy (WLE) as a control.

METHODS

Average-risk 50-75-year-old individuals who underwent screening colonoscopy at five referral centers were randomized to SD-CADe, CM-CADe, or WLE groups (1:1:1 ratio). Trainees and staff with an adenoma detection rate (ADR) of ≥35% were recruited. The primary outcome was ADR. Secondary outcomes were the proximal adenoma detection rate (pADR), advanced adenoma detection rate (AADR), and the number of adenomas, proximal adenomas, and advanced adenomas per colonoscopy (APC, pAPC, and AAPC, respectively).

RESULTS

The study enrolled 1200 participants. The ADR in the control, CM-CADe, and SD-CADe groups was 38.3%, 50.0%, and 54.8%, respectively. The pADR was 23.0%, 32.3%, and 38.8%, respectively. AADR was 6.0%, 10.3%, and 9.5%, respectively. After adjustment, the ADR and pADR in both intervention groups were significantly higher than in controls (all P<0.05). The APC in the control, CM-CADe, and SD-CADe groups was 0.66, 1.04, and 1.16, respectively. The pAPC was 0.33, 0.53, and 0.64, respectively, and the AAPC was 0.07, 0.12, and 0.10, respectively. Both CADe systems showed significantly higher APC and pAPC than WLE. AADR and AAPC were improved in both CADe groups versus control, although the differences were not statistically significant.

CONCLUSION

Even in high adenoma detectors, CADe significantly improved ADR and APC. The AADR tended to be higher with both systems, and this may enhance colorectal cancer prevention.

Performance comparison between two computer-aided detection colonoscopy models by trainees using different false positive thresholds: a cross-sectional study in Thailand

BACKGROUND/AIMS

This study aims to compare polyp detection performance of "Deep-GI," a newly developed artificial intelligence (AI) model, to a previously validated AI model computer-aided polyp detection (CADe) using various false positive (FP) thresholds and determining the best threshold for each model.

METHODS

Colonoscopy videos were collected prospectively and reviewed by three expert endoscopists (gold standard), trainees, CADe (CAD EYE; Fujifilm Corp.), and Deep-GI. Polyp detection sensitivity (PDS), polyp miss rates (PMR), and false-positive alarm rates (FPR) were compared among the three groups using different FP thresholds for the duration of bounding boxes appearing on the screen.

RESULTS

In total, 170 colonoscopy videos were used in this study. Deep-GI showed the highest PDS (99.4% vs. 85.4% vs. 66.7%, p<0.01) and the lowest PMR (0.6% vs. 14.6% vs. 33.3%, p<0.01) when compared to CADe and trainees, respectively. Compared to CADe, Deep-GI demonstrated lower FPR at FP thresholds of ≥0.5 (12.1 vs. 22.4) and ≥1 second (4.4 vs. 6.8) (both p<0.05). However, when the threshold was raised to ≥1.5 seconds, the FPR became comparable (2 vs. 2.4, p=0.3), while the PMR increased from 2% to 10%.

CONCLUSION

Compared to CADe, Deep-GI demonstrated a higher PDS with significantly lower FPR at ≥0.5- and ≥1-second thresholds. At the ≥1.5-second threshold, both systems showed comparable FPR with increased PMR.

Residual effects of combined vibratory and plantar stimulation while seated influences plantar pressure and spatiotemporal gait measures in individuals with Parkinson's disease exhibiting freezing of gait

Introduction

Combined plantar pressure and vibratory stimulation has been shown to decrease freezing of gait (FOG) episodes and improve spatiotemporal gait parameters compared to single stimulation in Parkinson's disease (PD) patients with FOG. However, the effect of combined plantar stimulations on plantar pressure analysis has never been explored.

Methods

Forty PD patients with frequent FOG were allocated to either FOG shoes embedded with a 100 Hz vibratory stimulation at the Achilles tendons and a soft thickened silicone pad at the hallux and sole, or sham shoes with a non-working vibratory motor and a flat non-pressure silicone pad (20 patients per arm) while seated for 96 s. The objective gait and plantar pressure analysis were measured immediately after the stimulation. Outcomes included the normalized percentage of changes in percent FOG (%FOG) and plantar pressure in the heel-strike and push-off phase that were compared between pre- and post-stimulations.

Results

The FOG shoes group showed significantly decreased %FOG (81.5 ± 28.9% vs. 6.8 ± 22.1%, p < 0.001), plantar pressure in the heel-strike (47.8 ± 43.7% vs. 4.3 ± 9.8%, p < 0.001), plantar pressure in the push-off (57.7 ± 59.6% vs. 6.2 ± 11.6%, p < 0.001), force time integral (FTI) (40.9 ± 32.5% vs. 6.6 ± 17.3%, p < 0.001), and decreased heel contact time (19.3 ± 12.3% vs. 22.7 ± 32.5%, p < 0.001) when compared to the sham group. There was a strong negative correlation between %FOG and peak plantar pressure (r = -0.440, p = 0.005), plantar pressure in the heel-strike (r = -0.847, p < 0.001).

Conclusion

Our study demonstrated that the FOG shoe could decrease FOG episodes by improving the heel-strike pressure, toe push-off and normalized heel-to-toe plantar pressure, suggesting that modification inputs from the peripheral sensory systems might significant improvement in FOG in PD.

Explainable Depression Detection Based on Facial Expression Using LSTM on Attentional Intermediate Feature Fusion with Label Smoothing

Machine learning is used for a fast pre-diagnosis approach to prevent the effects of Major Depressive Disorder (MDD). The objective of this research is to detect depression using a set of important facial features extracted from interview video, e.g., radians, gaze at angles, action unit intensity, etc. The model is based on LSTM with an attention mechanism. It aims to combine those features using the intermediate fusion approach. The label smoothing was presented to further improve the model's performance. Unlike other black-box models, the integrated gradient was presented as the model explanation to show important features of each patient. The experiment was conducted on 474 video samples collected at Chulalongkorn University. The data set was divided into 134 depressed and 340 non-depressed categories. The results showed that our model is the winner, with a 88.89% F1-score, 87.03% recall, 91.67% accuracy, and 91.40% precision. Moreover, the model can capture important features of depression, including head turning, no specific gaze, slow eye movement, no smiles, frowning, grumbling, and scowling, which express a lack of concentration, social disinterest, and negative feelings that are consistent with the assumptions in the depressive theories.

Real-time gastric intestinal metaplasia diagnosis tailored for bias and noisy-labeled data with multiple endoscopic imaging

This work presents real-time segmentation viz. gastric intestinal metaplasia (GIM). Recently, GIM segmentation of endoscopic images has been carried out to differentiate GIM from a healthy stomach. However, real-time detection is difficult to achieve. Conditions are challenging, and include multiple color modes (white light endoscopy and narrow-band imaging), other abnormal lesions (erosion and ulcer), noisy labels etc. Herein, our model is based on BiSeNet and can overcome the many issues regarding GIM. Application of auxiliary head and additional loss are seen to improve performance as well as enhance multiple color modes accurately. Further, multiple pre-processing techniques are utilized for leveraging detection performance: namely, location-wise negative sampling, jigsaw augmentation, and label smoothing. Finally, the decision threshold can be adjusted separately for each color mode. Work undertaken at King Chulalongkorn Memorial Hospital examined 940 histologically proven GIM images and 1239 non-GIM images, obtained over 173 frames per second (FPS). In terms of accuracy, our model is seen to outperform all baselines. Our results demonstrate sensitivity, specificity, positive predictive, negative predictive, accuracy, and mean intersection over union (IoU), achieving GIM segmentation values of 91%, 96%, 91%, 91%, 96%, and 55%, respectively.

Vibratory and plantar pressure stimulation: Steps to improve freezing of gait in Parkinson's disease

INTRODUCTION

Freezing of gait (FOG) is a devastating symptom that develops in patients with advanced Parkinson's disease (PD) and is often unresponsive to pharmacological treatment. Recent research suggests that FOG may result from dysfunctional plantar peripheral sensory systems. The impact of combined plantar pressure and vibratory stimulation over vibratory or pressure alone on FOG remains unexplored.

METHODS

PD patients with FOG were randomised into four groups and treated with combined vibratory and pressure stimulation, vibratory stimulation alone, pressure stimulation alone, or controls (no stimulation). Vibratory stimulation targeted both Achilles' tendons. Simultaneous bilateral pressure stimulation was applied to the first hallux, first metatarsal bone, and the sole. The primary outcome included normalized percent changes in percent FOG measured both pre- and immediately post-stimulation. Other outcomes including clinical rating scale, response to questionnaires, number and duration of freezing episodes, and spatiotemporal gait parameters at pre- and freezing episodes were also explored.

RESULTS

Sixty PD patients participated in the study. Patients who were treated with combined vibratory and pressure stimulation responded with significant decreases in normalized percent changes of percent FOG (62.75 ± 25.54%, p < 0.001) compared with those treated with vibration alone (11.38 ± 8.29%, p < 0.001), pressure alone (15.15 ± 16.18%, p < 0.001), or controls (8.59 ± 16.85%, p < 0.001).

CONCLUSION

Our study demonstrated the benefit of combined vibratory and pressure stimulation on FOG suggesting that this strategy might be developed as a novel treatment modality for PD patients with FOG.

Real-time semantic segmentation of gastric intestinal metaplasia using a deep learning approach

Background/Aims

Previous artificial intelligence (AI) models attempting to segment gastric intestinal metaplasia (GIM) areas have failed to be deployed in real-time endoscopy due to their slow inference speeds. Here, we propose a new GIM segmentation AI model with inference speeds faster than 25 frames per second that maintains a high level of accuracy.

Methods

Investigators from Chulalongkorn University obtained 802 histological-proven GIM images for AI model training. Four strategies were proposed to improve the model accuracy. First, transfer learning was employed to the public colon datasets. Second, an image preprocessing technique contrast-limited adaptive histogram equalization was employed to produce clearer GIM areas. Third, data augmentation was applied for a more robust model. Lastly, the bilateral segmentation network model was applied to segment GIM areas in real time. The results were analyzed using different validity values.

Results

From the internal test, our AI model achieved an inference speed of 31.53 frames per second. GIM detection showed sensitivity, specificity, positive predictive, negative predictive, accuracy, and mean intersection over union in GIM segmentation values of 93%, 80%, 82%, 92%, 87%, and 57%, respectively.

Conclusions

The bilateral segmentation network combined with transfer learning, contrast-limited adaptive histogram equalization, and data augmentation can provide high sensitivity and good accuracy for GIM detection and segmentation.

Federated learning for predicting clinical outcomes in patients with COVID-19

Federated learning (FL) is a method used for training artificial intelligence models with data from multiple sources while maintaining data anonymity, thus removing many barriers to data sharing. Here we used data from 20 institutes across the globe to train a FL model, called EXAM (electronic medical record (EMR) chest X-ray AI model), that predicts the future oxygen requirements of symptomatic patients with COVID-19 using inputs of vital signs, laboratory data and chest X-rays. EXAM achieved an average area under the curve (AUC) >0.92 for predicting outcomes at 24 and 72 h from the time of initial presentation to the emergency room, and it provided 16% improvement in average AUC measured across all participating sites and an average increase in generalizability of 38% when compared with models trained at a single site using that site's data. For prediction of mechanical ventilation treatment or death at 24 h at the largest independent test site, EXAM achieved a sensitivity of 0.950 and specificity of 0.882. In this study, FL facilitated rapid data science collaboration without data exchange and generated a model that generalized across heterogeneous, unharmonized datasets for prediction of clinical outcomes in patients with COVID-19, setting the stage for the broader use of FL in healthcare.

Using Machine Learning for Predicting the Best Outcomes With Electrical Muscle Stimulation for Tremors in Parkinson's Disease

Recent studies have identified that peripheral stimulation in Parkinson’s disease (PD) is effective in tremor reduction, indicating that a peripheral feedback loop plays an important role in the tremor reset mechanism. This was an open-label, quasi-experimental, pre- and post-test design, single-blind, single-group study involving 20 tremor-dominant PD patients. The objective of this study is to explore the effect of electrical muscle stimulation (EMS) as an adjunctive treatment for resting tremor during “on” period and to identify the best machine learning model to predict the suitable stimulation level that will yield the longest period of tremor reduction or tremor reset time. In this study, we used a Parkinson’s glove to evaluate, stimulate, and quantify the tremors of PD patients. This adjustable glove incorporates a 3-axis gyroscope to measure tremor signals and an EMS to provide an on-demand muscle stimulation to suppress tremors. Machine learning models were applied to identify the suitable pulse amplitude (stimulation level) in five classes that led to the longest tremor reset time. The study was registered at the www.clinicaltrials.gov under the name “The Study of Rest Tremor Suppression by Using Electrical Muscle Stimulation” (NCT02370108). Twenty tremor-dominant PD patients were recruited. After applying an average pulse amplitude of 6.25 (SD 2.84) mA and stimulation period of 440.7 (SD 560.82) seconds, the total time of tremor reduction, or tremor reset time, was 329.90 (SD 340.91) seconds. A significant reduction in tremor parameters during stimulation was demonstrated by a reduction of Unified Parkinson’s Disease Rating Scale (UPDRS) scores, and objectively, with a reduction of gyroscopic data (p < 0.05, each). None of the subjects reported any serious adverse events. We also compared gyroscopic data with five machine learning techniques: Logistic Regression, Random Forest, Support Vector Machine (SVM), Neural Network (NN), and Long-Short-Term-Memory (LSTM). The machine learning model that gave the highest accuracy was LSTM, which obtained: accuracy = 0.865 and macro-F1 = 0.736. This study confirms the efficacy of EMS in the reduction of resting tremors in PD. LSTM was identified as the most effective model for predicting pulse amplitude that would elicit the longest tremor reset time. Our study provides further insight on the tremor reset mechanism in PD.

Federated Learning used for predicting outcomes in SARS-COV-2 patients

'Federated Learning' (FL) is a method to train Artificial Intelligence (AI) models with data from multiple sources while maintaining anonymity of the data thus removing many barriers to data sharing. During the SARS-COV-2 pandemic, 20 institutes collaborated on a healthcare FL study to predict future oxygen requirements of infected patients using inputs of vital signs, laboratory data, and chest x-rays, constituting the "EXAM" (EMR CXR AI Model) model. EXAM achieved an average Area Under the Curve (AUC) of over 0.92, an average improvement of 16%, and a 38% increase in generalisability over local models. The FL paradigm was successfully applied to facilitate a rapid data science collaboration without data exchange, resulting in a model that generalised across heterogeneous, unharmonized datasets. This provided the broader healthcare community with a validated model to respond to COVID-19 challenges, as well as set the stage for broader use of FL in healthcare.

Adversarial Evaluation of Robust Neural Sequential Tagging Methods for Thai Language

Sequential tagging tasks, such as Part-Of-Speech (POS) tagging and Named-Entity Recognition, are the building blocks of many natural language processing applications. Although prior works have reported promising results in standard settings, they often underperform on non-standard text, such as microblogs and social media. In this article, we introduce an adversarial evaluation scheme for the Thai language by creating adversarial examples based on known spelling errors. Furthermore, we propose novel methods including UNK masking, condition initialization with affixation embeddings, and untied-directional self-attention mechanism to enhance robustness and interpretability of the neural networks. We conducted experiments on two Thai corpora: BEST2010 and ORCHID. Our adversarial evaluation schemes reveal that bidirectional LSTM (BiLSTM) do not perform well on adversarial examples. Our best methods match the performance of the BiLSTM baseline model and outperform it on adversarial examples.

Semi-supervised Thai Sentence Segmentation Using Local and Distant Word Representations

A sentence is typically treated as the minimal syntactic unit used for extracting valuable information from a longer piece of text. However, in written Thai, there are no explicit sentence markers. We proposed a deep learning model for the task of sentence segmentation that includes three main contributions. First, we integrate n-gram embedding as a local representation to capture word groups near sentence boundaries. Second, to focus on the keywords of dependent clauses, we combine the model with a distant representation obtained from self-attention modules. Finally, due to the scarcity of labeled data, for which annotation is difficult and time-consuming, we also investigate and adapt Cross-View Training (CVT) as a semi-supervised learning technique, allowing us to utilize unlabeled data to improve the model representations. In the Thai sentence segmentation experiments, our model reduced the relative error by 7.4% and 10.5% compared with the baseline models on the Orchid and UGWC datasets, respectively. We also applied our model to the task of pronunciation recovery on the IWSLT English dataset. Our model outperformed the prior sequence tagging models, achieving a relative error reduction of 2.5%. Ablation studies revealed that utilizing n-gram presentations was the main contributing factor for Thai, while the semi-supervised training helped the most for English.

Relation extraction between bacteria and biotopes from biomedical texts with attention mechanisms and domain-specific contextual representations

BACKGROUND

The Bacteria Biotope (BB) task is a biomedical relation extraction (RE) that aims to study the interaction between bacteria and their locations. This task is considered to pertain to fundamental knowledge in applied microbiology. Some previous investigations conducted the study by applying feature-based models; others have presented deep-learning-based models such as convolutional and recurrent neural networks used with the shortest dependency paths (SDPs). Although SDPs contain valuable and concise information, some parts of crucial information that is required to define bacterial location relationships are often neglected. Moreover, the traditional word-embedding used in previous studies may suffer from word ambiguation across linguistic contexts.

RESULTS

Here, we present a deep learning model for biomedical RE. The model incorporates feature combinations of SDPs and full sentences with various attention mechanisms. We also used pre-trained contextual representations based on domain-specific vocabularies. To assess the model's robustness, we introduced a mean F1 score on many models using different random seeds. The experiments were conducted on the standard BB corpus in BioNLP-ST'16. Our experimental results revealed that the model performed better (in terms of both maximum and average F1 scores; 60.77% and 57.63%, respectively) compared with other existing models.

CONCLUSIONS

We demonstrated that our proposed contributions to this task can be used to extract rich lexical, syntactic, and semantic features that effectively boost the model's performance. Moreover, we analyzed the trade-off between precision and recall to choose the proper cut-off to use in real-world applications.