University of Leicester

We are pleased to introduce the collection Frontiers in Medical Technology – Highlights in Medtech Data Analytics 2021/22. This collection will welcome and showcase a selection of high impact articles authored by leaders in the field. The work presented here highlights the broad diversity of research performed across the Medtech Data Analytics section, and aims to put a spotlight on the main areas of interest. This collection aims to further support Frontiers’ strong community by shining a spotlight on our authors' highly impactful research.

OOI CHIA YEE

Nanjing Medical University

University of Granada

Views

Article Views

Article Downloads

Frontiers in Medical Technology

Medtech Data Analytics

Highlights in Medtech Data Analytics 2021/22

BackgroundIn Kampo medicine, tongue examination is used to diagnose the pathological condition “Sho,” but an objective evaluation method for its diagnostic ability has not been established. We constructed a tongue diagnosis electronic learning and evaluation system based on a standardized tongue image database.</sec>PurposeThis study aims to verify the practicality of this assessment system by evaluating the tongue diagnosis ability of Kampo specialists (KSs), medical professionals, and students.</sec>MethodsIn the first study, we analyzed the answer data of 15 KSs in an 80-question tongue diagnosis test that assesses eight aspects of tongue findings and evaluated the (i) test score, (ii) test difficulty and discrimination index, (iii) diagnostic consistency, and (iv) diagnostic match rate between KSs. In the second study, we administered a 20-question common Kampo test and analyzed the answer data of 107 medical professionals and 56 students that assessed the tongue color discrimination ability and evaluated the (v) correct answer rate, (vi) test difficulty, and (vii) factors related to the correct answer rate.</sec>ResultIn the first study, the average test score was 62.2 ± 10.7 points. Twenty-eight questions were difficult (correct answer rate, <50%), 34 were moderate (50%–85%), and 18 were easy (≥85%). Regarding intrarater reliability, the average diagnostic match rate of five KSs involved in database construction was 0.66 ± 0.08, and as for interrater reliability, the diagnostic match rate between the 15 KSs was 0.52 (95% confidence interval, 0.38–0.65) for Gwet's agreement coefficient 1, and the degree of the match rate was moderate. In the second study, the difficulty level of questions was moderate, with a correct rate of 81.3% for medical professionals and 82.1% for students. The discrimination index was good for medical professionals (0.35) and poor for students (0.06). Among medical professionals, the correct answer group of this question had a significantly higher total score on the Kampo common test than the incorrect answer group (85.3 ± 8.4 points vs. 75.8 ± 11.8 points, p < 0.01).</sec>ConclusionThis system can objectively evaluate tongue diagnosis ability and has high practicality. Utilizing this system can be expected to contribute to improving learners’ tongue diagnosis ability and standardization of tongue diagnosis.</sec>

Background: In Kampo medicine, tongue examination is used to diagnose the pathological condition “Sho,” but an objective evaluation method for its diagnostic ability has not been established. We constructed a tongue diagnosis electronic learning and evaluation system based on a standardized tongue image database.Purpose: This study aims to verify the practicality of this assessment system by evaluating the tongue diagnosis ability of Kampo specialists (KSs), medical professionals, and students.Methods: In the first study, we analyzed the answer data of 15 KSs in an 80-question tongue diagnosis test that assesses eight aspects of tongue findings and evaluated the (i) test score, (ii) test difficulty and discrimination index, (iii) diagnostic consistency, and (iv) diagnostic match rate between KSs. In the second study, we administered a 20-question common Kampo test and analyzed the answer data of 107 medical professionals and 56 students that assessed the tongue color discrimination ability and evaluated the (v) correct answer rate, (vi) test difficulty, and (vii) factors related to the correct answer rate.Result: In the first study, the average test score was 62.2 ± 10.7 points. Twenty-eight questions were difficult (correct answer rate, &lt;50%), 34 were moderate (50%–85%), and 18 were easy (≥85%). Regarding intrarater reliability, the average diagnostic match rate of five KSs involved in database construction was 0.66 ± 0.08, and as for interrater reliability, the diagnostic match rate between the 15 KSs was 0.52 (95% confidence interval, 0.38–0.65) for Gwet's agreement coefficient 1, and the degree of the match rate was moderate. In the second study, the difficulty level of questions was moderate, with a correct rate of 81.3% for medical professionals and 82.1% for students. The discrimination index was good for medical professionals (0.35) and poor for students (0.06). Among medical professionals, the correct answer group of this question had a significantly higher total score on the Kampo common test than the incorrect answer group (85.3 ± 8.4 points vs. 75.8 ± 11.8 points, p &lt; 0.01).Conclusion: This system can objectively evaluate tongue diagnosis ability and has high practicality. Utilizing this system can be expected to contribute to improving learners’ tongue diagnosis ability and standardization of tongue diagnosis.

Objective evaluation of tongue diagnosis ability using a tongue diagnosis e-learning/e-assessment system based on a standardized tongue image database

It is often assumed that healthy people have the genuine ability to maintain tight blood glucose regulation. However, a few recent studies revealed that glucose dysregulation such as hyperglycemia may occur even in people who are considered normoglycemic by standard measures and were more prevalent than initially thought, suggesting that more investigations are needed to fully understand the within-day glucose dynamics of healthy people. In this paper, we conducted an analysis on a multi-modal dataset to examine the relationships between glycemic variability when people were awake and that when they were sleeping. The interstitial glucose levels were measured with a wearable continuous glucose monitoring (CGM) technology FreeStyle Libre 2 at every 15 min interval. In contrast to the traditional single-time-point measurements, the CGM data allow the investigation into the temporal patterns of glucose dynamics at high granularity. Sleep onset and offset timestamps were recorded daily with a Fitbit Charge 3 wristband. Our analysis leveraged the sleep data to split the glucose readings into segments of awake-time and in-sleep, instead of using fixed cut-off time points as has been done in existing literature. We combined repeated measure correlation analysis and quantitative association rules mining, together with an original post-filtering method, to identify significant and most relevant associations. Our results showed that low overall glucose in awake time was strongly correlated to low glucose in subsequent sleep, which in turn correlated to overall low glucose in the next day. Moreover, both analysis techniques identified significant associations between the minimal glucose reading in sleep and the low blood glucose index the next day. In addition, the association rules discovered in this study achieved high confidence (0.75–0.88) and lift (4.1–11.5), which implies that the proposed post-filtering method was effective in selecting quality rules.

It is often assumed that healthy people have the genuine ability to maintain tight blood glucose regulation. However, a few recent studies revealed that glucose dysregulation such as hyperglycemia may occur even in people who are considered normoglycemic by standard measures and were more prevalent than initially thought, suggesting that more investigations are needed to fully understand the within-day glucose dynamics of healthy people. In this paper, we conducted an analysis on a multi-modal dataset to examine the relationships between glycemic variability when people were awake and that when they were sleeping. The interstitial glucose levels were measured with a wearable continuous glucose monitoring (CGM) technology FreeStyle Libre 2 at every 15&#x2005;min interval. In contrast to the traditional single-time-point measurements, the CGM data allow the investigation into the temporal patterns of glucose dynamics at high granularity. Sleep onset and offset timestamps were recorded daily with a Fitbit Charge 3 wristband. Our analysis leveraged the sleep data to split the glucose readings into segments of awake-time and in-sleep, instead of using fixed cut-off time points as has been done in existing literature. We combined repeated measure correlation analysis and quantitative association rules mining, together with an original post-filtering method, to identify significant and most relevant associations. Our results showed that low overall glucose in awake time was strongly correlated to low glucose in subsequent sleep, which in turn correlated to overall low glucose in the next day. Moreover, both analysis techniques identified significant associations between the minimal glucose reading in sleep and the low blood glucose index the next day. In addition, the association rules discovered in this study achieved high confidence (0.75&#x2013;0.88) and lift (4.1&#x2013;11.5), which implies that the proposed post-filtering method was effective in selecting quality rules.

Mining associations between glycemic variability in awake-time and in-sleep among non-diabetic adults

Deep neural networks (DNNs) have started to find their role in the modern healthcare system. DNNs are being developed for diagnosis, prognosis, treatment planning, and outcome prediction for various diseases. With the increasing number of applications of DNNs in modern healthcare, their trustworthiness and reliability are becoming increasingly important. An essential aspect of trustworthiness is detecting the performance degradation and failure of deployed DNNs in medical settings. The softmax output values produced by DNNs are not a calibrated measure of model confidence. Softmax probability numbers are generally higher than the actual model confidence. The model confidence-accuracy gap further increases for wrong predictions and noisy inputs. We employ recently proposed Bayesian deep neural networks (BDNNs) to learn uncertainty in the model parameters. These models simultaneously output the predictions and a measure of confidence in the predictions. By testing these models under various noisy conditions, we show that the (learned) predictive confidence is well calibrated. We use these reliable confidence values for monitoring performance degradation and failure detection in DNNs. We propose two different failure detection methods. In the first method, we define a fixed threshold value based on the behavior of the predictive confidence with changing signal-to-noise ratio (SNR) of the test dataset. The second method learns the threshold value with a neural network. The proposed failure detection mechanisms seamlessly abstain from making decisions when the confidence of the BDNN is below the defined threshold and hold the decision for manual review. Resultantly, the accuracy of the models improves on the unseen test samples. We tested our proposed approach on three medical imaging datasets: PathMNIST, DermaMNIST, and OrganAMNIST, under different levels and types of noise. An increase in the noise of the test images increases the number of abstained samples. BDNNs are inherently robust and show more than 10% accuracy improvement with the proposed failure detection methods. The increased number of abstained samples or an abrupt increase in the predictive variance indicates model performance degradation or possible failure. Our work has the potential to improve the trustworthiness of DNNs and enhance user confidence in the model predictions.

Deep neural networks (DNNs) have started to find their role in the modern healthcare system. DNNs are being developed for diagnosis, prognosis, treatment planning, and outcome prediction for various diseases. With the increasing number of applications of DNNs in modern healthcare, their trustworthiness and reliability are becoming increasingly important. An essential aspect of trustworthiness is detecting the performance degradation and failure of deployed DNNs in medical settings. The softmax output values produced by DNNs are not a calibrated measure of model confidence. Softmax probability numbers are generally higher than the actual model confidence. The model confidence-accuracy gap further increases for wrong predictions and noisy inputs. We employ recently proposed Bayesian deep neural networks (BDNNs) to learn uncertainty in the model parameters. These models simultaneously output the predictions and a measure of confidence in the predictions. By testing these models under various noisy conditions, we show that the (learned) predictive confidence is well calibrated. We use these reliable confidence values for monitoring performance degradation and failure detection in DNNs. We propose two different failure detection methods. In the first method, we define a fixed threshold value based on the behavior of the predictive confidence with changing signal-to-noise ratio (SNR) of the test dataset. The second method learns the threshold value with a neural network. The proposed failure detection mechanisms seamlessly abstain from making decisions when the confidence of the BDNN is below the defined threshold and hold the decision for manual review. Resultantly, the accuracy of the models improves on the unseen test samples. We tested our proposed approach on three medical imaging datasets: PathMNIST, DermaMNIST, and OrganAMNIST, under different levels and types of noise. An increase in the noise of the test images increases the number of abstained samples. BDNNs are inherently robust and show more than 10% accuracy improvement with the proposed failure detection methods. The increased number of abstained samples or an abrupt increase in the predictive variance indicates model performance degradation or possible failure. Our work has the potential to improve the trustworthiness of DNNs and enhance user confidence in the model predictions.

Failure Detection in Deep Neural Networks for Medical Imaging

We have studied brain connectivity using a biologically inspired in silico model of the visual pathway consisting of the lateral geniculate nucleus (LGN) of the thalamus, and layers 4 and 6 of the primary visual cortex. The connectivity parameters in the model are informed by the existing anatomical parameters from mammals and rodents. In the base state, the LGN and layer 6 populations in the model oscillate with dominant alpha frequency, while the layer 4 oscillates in the theta band. By changing intra-cortical hyperparameters, specifically inhibition from layer 6 to layer 4, we demonstrate a transition to alpha mode for all the populations. Furthermore, by increasing the feedforward connectivities in the thalamo-cortico-thalamic loop, we could transition into the beta band for all the populations. On looking closely, we observed that the origin of this beta band is in the layer 6 (infragranular layers); lesioning the thalamic feedback from layer 6 removed the beta from the LGN and the layer 4. This agrees with existing physiological studies where it is shown that beta rhythm is generated in the infragranular layers. Lastly, we present a case study to demonstrate a neurological condition in the model. By changing connectivities in the network, we could simulate the condition of significant (P < 0.001) decrease in beta band power and a simultaneous increase in the theta band power, similar to that observed in Schizophrenia patients. Overall, we have shown that the connectivity changes in a simple visual thalamocortical in silico model can simulate state changes in the brain corresponding to both health and disease conditions.

We have studied brain connectivity using a biologically inspired in silico model of the visual pathway consisting of the lateral geniculate nucleus (LGN) of the thalamus, and layers 4 and 6 of the primary visual cortex. The connectivity parameters in the model are informed by the existing anatomical parameters from mammals and rodents. In the base state, the LGN and layer 6 populations in the model oscillate with dominant alpha frequency, while the layer 4 oscillates in the theta band. By changing intra-cortical hyperparameters, specifically inhibition from layer 6 to layer 4, we demonstrate a transition to alpha mode for all the populations. Furthermore, by increasing the feedforward connectivities in the thalamo-cortico-thalamic loop, we could transition into the beta band for all the populations. On looking closely, we observed that the origin of this beta band is in the layer 6 (infragranular layers); lesioning the thalamic feedback from layer 6 removed the beta from the LGN and the layer 4. This agrees with existing physiological studies where it is shown that beta rhythm is generated in the infragranular layers. Lastly, we present a case study to demonstrate a neurological condition in the model. By changing connectivities in the network, we could simulate the condition of significant (P &lt; 0.001) decrease in beta band power and a simultaneous increase in the theta band power, similar to that observed in Schizophrenia patients. Overall, we have shown that the connectivity changes in a simple visual thalamocortical in silico model can simulate state changes in the brain corresponding to both health and disease conditions.

In silico Effects of Synaptic Connections in the Visual Thalamocortical Pathway

ObjectiveTo evaluate the experience of the physician of the technical success in high-intensity focused ultrasound (HIFU) ablation of uterine fibroids with a nonperfused volume ratio (NPVR) of at least 80%.</sec>MethodsPatients from a 20-center prospective study were enrolled in this study. In this study, among the 20 clinical centers, five centers had physician with >3 years of HIFU experience, and the other 15 centers initiated HIFU therapy <3 years, were defined as the experienced group and the inexperienced group, respectively. Technical success was defined as achieving NPVR ≥ 80% of uterine fibroids with no major complications and it was defined as the successful group; otherwise, it was defined as the unsuccessful group.</sec>ResultsA total of 1,352 patients were included at the age of 41.32 ± 5.08 years. The mean NPVR (87.48 ± 14.91%) was obtained in the inexperienced group (86.50 ± 15.76%) and in the experienced group (89.21 ± 13.12%), respectively. The multivariate analysis showed that the volume of uterus, location of fibroids, and physician experience were significantly correlated with technical success (p < 0.05). In the experienced group, 82.20% of uterine fibroids obtained NPVR ≥ 80%, compared with 75.32% in the inexperienced group, and the difference was significant (p = 0.003). The technical success rate of the experienced group was 82.00% which was higher than 75.20% of the inexperienced group (p = 0.004).</sec>ConclusionIn technical success of achieving NPVR ≥ 80%, experience of the physician was positively correlated with technical success; NPVR and major complications for the inexperienced group were comparable to those of the experienced group from a clinical perspective; inexperienced physicians could reach NPVR ≥ 80% of sufficient ablation and were trustworthy in efficacy. Smaller uterus and fibroids of anterior wall were correlated with better technical success; experienced physicians still have better technical success when choosing patients with larger uterus, contributing to clinical decision-making and patient referral.</sec>

Objective: To evaluate the experience of the physician of the technical success in high-intensity focused ultrasound (HIFU) ablation of uterine fibroids with a nonperfused volume ratio (NPVR) of at least 80%.Methods: Patients from a 20-center prospective study were enrolled in this study. In this study, among the 20 clinical centers, five centers had physician with &gt;3 years of HIFU experience, and the other 15 centers initiated HIFU therapy &lt;3 years, were defined as the experienced group and the inexperienced group, respectively. Technical success was defined as achieving NPVR ≥ 80% of uterine fibroids with no major complications and it was defined as the successful group; otherwise, it was defined as the unsuccessful group.Results: A total of 1,352 patients were included at the age of 41.32 ± 5.08 years. The mean NPVR (87.48 ± 14.91%) was obtained in the inexperienced group (86.50 ± 15.76%) and in the experienced group (89.21 ± 13.12%), respectively. The multivariate analysis showed that the volume of uterus, location of fibroids, and physician experience were significantly correlated with technical success (p &lt; 0.05). In the experienced group, 82.20% of uterine fibroids obtained NPVR ≥ 80%, compared with 75.32% in the inexperienced group, and the difference was significant (p = 0.003). The technical success rate of the experienced group was 82.00% which was higher than 75.20% of the inexperienced group (p = 0.004).Conclusion: In technical success of achieving NPVR ≥ 80%, experience of the physician was positively correlated with technical success; NPVR and major complications for the inexperienced group were comparable to those of the experienced group from a clinical perspective; inexperienced physicians could reach NPVR ≥ 80% of sufficient ablation and were trustworthy in efficacy. Smaller uterus and fibroids of anterior wall were correlated with better technical success; experienced physicians still have better technical success when choosing patients with larger uterus, contributing to clinical decision-making and patient referral.

Physician Experience in Technical Success of Achieving NPVR ≥ 80% of High-Intensity Focused Ultrasound Ablation for Uterine Fibroids: A Multicenter Study

Tongue examination is an important diagnostic method for judging pathological conditions in Kampo (traditional Japanese medicine), but it is not easy for beginners to learn the diagnostic technique. One reason is that there are few objective diagnostic criteria for tongue examination findings, and the educational method for tongue examination is not standardized in Japan, warranting the need for a tongue image database for e-learning systems that could dramatically improve the efficiency of education. Therefore, we constructed a database comprising tongue images whose findings were determined on the basis of votes given by five Kampo medicine specialists (KMSs) and confirmed the educational usefulness of the database for tongue diagnosis e-learning systems. The study was conducted in the following five steps: development of a tongue imaging collection system, collection of tongue images, evaluation and annotation of tongue images, development of a tongue diagnosis e-learning system, and verification of the educational usefulness of this system. Five KMSs evaluated the tongue images obtained from 125 participants in the following eight aspects: (i) tongue body size, (ii) tongue body color, (iii) tongue body dryness and wetness, (iv) tooth marks on the edge of the tongue, (v) cracks on the surface of the tongue, (vi) thickness of tongue coating, (vii) color of tongue coating, and (viii) dryness and wetness of tongue coating. Medical students (MSs) were given a tongue diagnosis test using an e-learning system after a lecture on tongue diagnosis. The cumulative and individual match rates (%) (individual match rates of 100% (5/5), 80% (4/5), and 60% (3/5) are shown in parentheses, respectively) were as follows: (i) tongue body size: 92.8 (26.4/26.4/40.0); (ii) tongue body color: 83.2 (10.4/20.8/52.0); (iii) tongue body dryness and wetness: 88.8 (13.6/34.4/40.8); (iv) tooth marks on the edge of the tongue: 88.8 (6.4/35.2/47.2); (v) cracks on the surface of the tongue: 96.8 (24.0/35.2/37.6); (vi) thickness of tongue coating: 84.8 (7.2/21.6/56.0); (vii) color of tongue coating: 88.0 (15.2/37.6/35.2); and (viii) dryness and wetness of tongue coating: 74.4 (4.8/19.2/50.4). The test showed that the tongue diagnosis ability of MSs who attended a lecture on tongue diagnosis was almost the same as that of KMSs. We successfully constructed a tongue image database standardized for training specialists on tongue diagnosis and confirmed the educational usefulness of the e-learning system using a database. This database will contribute to the standardization and popularization of Kampo education.

Tongue examination is an important diagnostic method for judging pathological conditions in Kampo (traditional Japanese medicine), but it is not easy for beginners to learn the diagnostic technique. One reason is that there are few objective diagnostic criteria for tongue examination findings, and the educational method for tongue examination is not standardized in Japan, warranting the need for a tongue image database for e-learning systems that could dramatically improve the efficiency of education. Therefore, we constructed a database comprising tongue images whose findings were determined on the basis of votes given by five Kampo medicine specialists (KMSs) and confirmed the educational usefulness of the database for tongue diagnosis e-learning systems. The study was conducted in the following five steps: development of a tongue imaging collection system, collection of tongue images, evaluation and annotation of tongue images, development of a tongue diagnosis e-learning system, and verification of the educational usefulness of this system. Five KMSs evaluated the tongue images obtained from 125 participants in the following eight aspects: (i) tongue body size, (ii) tongue body color, (iii) tongue body dryness and wetness, (iv) tooth marks on the edge of the tongue, (v) cracks on the surface of the tongue, (vi) thickness of tongue coating, (vii) color of tongue coating, and (viii) dryness and wetness of tongue coating. Medical students (MSs) were given a tongue diagnosis test using an e-learning system after a lecture on tongue diagnosis. The cumulative and individual match rates (%) (individual match rates of 100% (5/5), 80% (4/5), and 60% (3/5) are shown in parentheses, respectively) were as follows: (i) tongue body size: 92.8 (26.4/26.4/40.0); (ii) tongue body color: 83.2 (10.4/20.8/52.0); (iii) tongue body dryness and wetness: 88.8 (13.6/34.4/40.8); (iv) tooth marks on the edge of the tongue: 88.8 (6.4/35.2/47.2); (v) cracks on the surface of the tongue: 96.8 (24.0/35.2/37.6); (vi) thickness of tongue coating: 84.8 (7.2/21.6/56.0); (vii) color of tongue coating: 88.0 (15.2/37.6/35.2); and (viii) dryness and wetness of tongue coating: 74.4 (4.8/19.2/50.4). The test showed that the tongue diagnosis ability of MSs who attended a lecture on tongue diagnosis was almost the same as that of KMSs. We successfully constructed a tongue image database standardized for training specialists on tongue diagnosis and confirmed the educational usefulness of the e-learning system using a database. This database will contribute to the standardization and popularization of Kampo education.