Evaluating Deep Q-Learning Algorithms for Controlling Blood Glucose in In Silico Type 1 Diabetes.

Patients with type 1 diabetes must continually decide how much insulin to inject before each meal to maintain blood glucose levels within a healthy range. Recent research has worked on a solution for this burden, showing the potential of reinforcement learning as an emerging approach for the task of controlling blood glucose levels. In this paper, we test and evaluate several deep Q-learning algorithms for automated and personalized blood glucose regulation in an in silico type 1 diabetes patient with the goal of estimating and delivering proper insulin doses. The proposed algorithms are model-free approaches with no prior information about the patient. We used the Hovorka model with meal variation and carbohydrate counting errors to simulate the patient included in this work. Our experiments compare different deep Q-learning extensions showing promising results controlling blood glucose levels, with some of the proposed algorithms outperforming standard baseline treatment.

Artificial Intelligence Implementation in Healthcare: A Theory-Based Scoping Review of Barriers and Facilitators.

There is a large proliferation of complex data-driven artificial intelligence (AI) applications in many aspects of our daily lives, but their implementation in healthcare is still limited. This scoping review takes a theoretical approach to examine the barriers and facilitators based on empirical data from existing implementations. We searched the major databases of relevant scientific publications for articles related to AI in clinical settings, published between 2015 and 2021. Based on the theoretical constructs of the Consolidated Framework for Implementation Research (CFIR), we used a deductive, followed by an inductive, approach to extract facilitators and barriers. After screening 2784 studies, 19 studies were included in this review. Most of the cited facilitators were related to engagement with and management of the implementation process, while the most cited barriers dealt with the intervention's generalizability and interoperability with existing systems, as well as the inner settings' data quality and availability. We noted per-study imbalances related to the reporting of the theoretic domains. Our findings suggest a greater need for implementation science expertise in AI implementation projects, to improve both the implementation process and the quality of scientific reporting.

Automatic detection of the mental foramen for estimating mandibular cortical width in dental panoramic radiographs: the seventh survey of the Tromsø Study (Tromsø7) in 2015-2016.

OBJECTIVE: To apply deep learning to a data set of dental panoramic radiographs to detect the mental foramen for automatic assessment of the mandibular cortical width. METHODS: Data from the seventh survey of the Tromsø Study (Tromsø7) were used. The data set contained 5197 randomly chosen dental panoramic radiographs. Four pretrained object detectors were tested. We randomly chose 80% of the data for training and 20% for testing. Models were trained using GeForce RTX 2080 Ti with 11 GB GPU memory (NVIDIA Corporation, Santa Clara, CA, USA). Python programming language version 3.7 was used for analysis. RESULTS: The EfficientDet-D0 model showed the highest average precision of 0.30. When the threshold to regard a prediction as correct (intersection over union) was set to 0.5, the average precision was 0.79. The RetinaNet model achieved the lowest average precision of 0.23, and the precision was 0.64 when the intersection over union was set to 0.5. The procedure to estimate mandibular cortical width showed acceptable results. Of 100 random images, the algorithm produced an output 93 times, 20 of which were not visually satisfactory. CONCLUSIONS: EfficientDet-D0 effectively detected the mental foramen. Methods for estimating bone quality are important in radiology and require further development.

Curve-Based Classification Approach for Hyperspectral Dermatologic Data Processing.

This paper shows new contributions in the detection of skin cancer, where we present the use of a customized hyperspectral system that captures images in the spectral range from 450 to 950 nm. By choosing a 7 × 7 sub-image of each channel in the hyperspectral image (HSI) and then taking the mean and standard deviation of these sub-images, we were able to make fits of the resulting curves. These fitted curves had certain characteristics, which then served as a basis of classification. The most distinct fit was for the melanoma pigmented skin lesions (PSLs), which is also the most aggressive malignant cancer. Furthermore, we were able to classify the other PSLs in malignant and benign classes. This gives us a rather complete classification method for PSLs with a novel perspective of the classification procedure by exploiting the variability of each channel in the HSI.

Soft thresholding schemes for multiple signal classification algorithm.

Multiple signal classification algorithm (MUSICAL) exploits temporal fluctuations in fluorescence intensity to perform super-resolution microscopy by computing the value of a super-resolving indicator function across a fine sample grid. A key step in the algorithm is the separation of the measurements into signal and noise subspaces, based on a single user-specified parameter called the threshold. The resulting image is strongly sensitive to this parameter and the subjectivity arising from multiple practical factors makes it difficult to determine the right rule of selection. We address this issue by proposing soft thresholding schemes derived from a new generalized framework for indicator function design. We show that the new schemes significantly alleviate the subjectivity and sensitivity of hard thresholding while retaining the super-resolution ability. We also evaluate the trade-off between resolution and contrast and the out-of-focus light rejection using the various indicator functions. Through this, we create significant new insights into the use and further optimization of MUSICAL for a wide range of practical scenarios.

Hyperspectral imaging and deep learning for the detection of breast cancer cells in digitized histological images.

In recent years, hyperspectral imaging (HSI) has been shown as a promising imaging modality to assist pathologists in the diagnosis of histological samples. In this work, we present the use of HSI for discriminating between normal and tumor breast cancer cells. Our customized HSI system includes a hyperspectral (HS) push-broom camera, which is attached to a standard microscope, and home-made software system for the control of image acquisition. Our HS microscopic system works in the visible and near-infrared (VNIR) spectral range (400 - 1000 nm). Using this system, 112 HS images were captured from histologic samples of human patients using 20× magnification. Cell-level annotations were made by an expert pathologist in digitized slides and were then registered with the HS images. A deep learning neural network was developed for the HS image classification, which consists of nine 2D convolutional layers. Different experiments were designed to split the data into training, validation and testing sets. In all experiments, the training and the testing set correspond to independent patients. The results show an area under the curve (AUCs) of more than 0.89 for all the experiments. The combination of HSI and deep learning techniques can provide a useful tool to aid pathologists in the automatic detection of cancer cells on digitized pathologic images.

De-Identifying Swedish EHR Text Using Public Resources in the General Domain.

Sensitive data is normally required to develop rule-based or train machine learning-based models for de-identifying electronic health record (EHR) clinical notes; and this presents important problems for patient privacy. In this study, we add non-sensitive public datasets to EHR training data; (i) scientific medical text and (ii) Wikipedia word vectors. The data, all in Swedish, is used to train a deep learning model using recurrent neural networks. Tests on pseudonymized Swedish EHR clinical notes showed improved precision and recall from 55.62% and 80.02% with the base EHR embedding layer, to 85.01% and 87.15% when Wikipedia word vectors are added. These results suggest that non-sensitive text from the general domain can be used to train robust models for de-identifying Swedish clinical text; and this could be useful in cases where the data is both sensitive and in low-resource languages.

Reinforcement learning application in diabetes blood glucose control: A systematic review.

BACKGROUND: Reinforcement learning (RL) is a computational approach to understanding and automating goal-directed learning and decision-making. It is designed for problems which include a learning agent interacting with its environment to achieve a goal. For example, blood glucose (BG) control in diabetes mellitus (DM), where the learning agent and its environment are the controller and the body of the patient respectively. RL algorithms could be used to design a fully closed-loop controller, providing a truly personalized insulin dosage regimen based exclusively on the patient's own data. OBJECTIVE: In this review we aim to evaluate state-of-the-art RL approaches to designing BG control algorithms in DM patients, reporting successfully implemented RL algorithms in closed-loop, insulin infusion, decision support and personalized feedback in the context of DM. METHODS: An exhaustive literature search was performed using different online databases, analyzing the literature from 1990 to 2019. In a first stage, a set of selection criteria were established in order to select the most relevant papers according to the title, keywords and abstract. Research questions were established and answered in a second stage, using the information extracted from the articles selected during the preliminary selection. RESULTS: The initial search using title, keywords, and abstracts resulted in a total of 404 articles. After removal of duplicates from the record, 347 articles remained. An independent analysis and screening of the records against our inclusion and exclusion criteria defined in Methods section resulted in removal of 296 articles, leaving 51 relevant articles. A full-text assessment was conducted on the remaining relevant articles, which resulted in 29 relevant articles that were critically analyzed. The inter-rater agreement was measured using Cohen Kappa test, and disagreements were resolved through discussion. CONCLUSIONS: The advances in health technologies and mobile devices have facilitated the implementation of RL algorithms for optimal glycemic regulation in diabetes. However, there exists few articles in the literature focused on the application of these algorithms to the BG regulation problem. Moreover, such algorithms are designed for control tasks as BG adjustment and their use have increased recently in the diabetes research area, therefore we foresee RL algorithms will be used more frequently for BG control in the coming years. Furthermore, in the literature there is a lack of focus on aspects that influence BG level such as meal intakes and physical activity (PA), which should be included in the control problem. Finally, there exists a need to perform clinical validation of the algorithms.

Hyperspectral Imaging for the Detection of Glioblastoma Tumor Cells in H&E Slides Using Convolutional Neural Networks.

Hyperspectral imaging (HSI) technology has demonstrated potential to provide useful information about the chemical composition of tissue and its morphological features in a single image modality. Deep learning (DL) techniques have demonstrated the ability of automatic feature extraction from data for a successful classification. In this study, we exploit HSI and DL for the automatic differentiation of glioblastoma (GB) and non-tumor tissue on hematoxylin and eosin (H&E) stained histological slides of human brain tissue. GB detection is a challenging application, showing high heterogeneity in the cellular morphology across different patients. We employed an HSI microscope, with a spectral range from 400 to 1000 nm, to collect 517 HS cubes from 13 GB patients using 20× magnification. Using a convolutional neural network (CNN), we were able to automatically detect GB within the pathological slides, achieving average sensitivity and specificity values of 88% and 77%, respectively, representing an improvement of 7% and 8% respectively, as compared to the results obtained using RGB (red, green, and blue) images. This study demonstrates that the combination of hyperspectral microscopic imaging and deep learning is a promising tool for future computational pathologies.

The effects of terlipressin and direct portacaval shunting on liver hemodynamics following 80% hepatectomy in the pig.

Liver failure is the major cause of death following liver resection. Post-resection portal venous pressure (PVP) predicts liver failure, is implicated in its pathogenesis, and when PVP is reduced, rates of liver dysfunction decrease. The aim of the present study was to characterize the hemodynamic, biochemical, and histological changes induced by 80% hepatectomy in non-cirrhotic pigs and determine if terlipressin or direct portacaval shunting can modulate these effects. Pigs were randomized (n=8/group) to undergo 80% hepatectomy alone (control); terlipressin (2 mg bolus + 0.5-1 mg/h) + 80% hepatectomy; or portacaval shunt (PCS) + 80% hepatectomy, and were maintained under terminal anesthesia for 8 h. The primary outcome was changed in PVP. Secondary outcomes included portal venous flow (PVF), hepatic arterial flow (HAF), and biochemical and histological markers of liver injury. Hepatectomy increased PVP (9.3 ± 0.4 mmHg pre-hepatectomy compared with 13.0 ± 0.8 mmHg post-hepatectomy, P<0.0001) and PVF/g liver (1.2 ± 0.2 compared with 6.0 ± 0.6 ml/min/g, P<0.0001) and decreased HAF (70.8 ± 5.0 compared with 41.8 ± 5.7 ml/min, P=0.002). Terlipressin and PCS reduced PVP (terlipressin = 10.4 ± 0.8 mmHg, P=0.046 and PCS = 8.3 ± 1.2 mmHg, P=0.025) and PVF (control = 869.0 ± 36.1 ml/min compared with terlipressin = 565.6 ± 25.7 ml/min, P<0.0001 and PCS = 488.4 ± 106.4 ml/min, P=0.002) compared with control. Treatment with terlipressin increased HAF (73.2 ± 11.3 ml/min) compared with control (40.3 ± 6.3 ml/min, P=0.026). The results of the present study suggest that terlipressin and PCS may have a role in the prevention and treatment of post-resection liver failure.

A novel scale-space approach for multinormality testing and the k-sample problem in the high dimension low sample size scenario.

Two classical multivariate statistical problems, testing of multivariate normality and the k-sample problem, are explored by a novel analysis on several resolutions simultaneously. The presented methods do not invert any estimated covariance matrix. Thereby, the methods work in the High Dimension Low Sample Size situation, i.e. when n ≤ p. The output, a significance map, is produced by doing a one-dimensional test for all possible resolution/position pairs. The significance map shows for which resolution/position pairs the null hypothesis is rejected. For the testing of multinormality, the Anderson-Darling test is utilized to detect potential departures from multinormality at different combinations of resolutions and positions. In the k-sample case, it is tested whether k data sets can be said to originate from the same unspecified discrete or continuous multivariate distribution. This is done by testing the k vectors corresponding to the same resolution/position pair of the k different data sets through the k-sample Anderson-Darling test. Successful demonstrations of the new methodology on artificial and real data sets are presented, and a feature selection scheme is demonstrated.

Control of Blood Glucose for Type-1 Diabetes by Using Reinforcement Learning with Feedforward Algorithm.

BACKGROUND: Type-1 diabetes is a condition caused by the lack of insulin hormone, which leads to an excessive increase in blood glucose level. The glucose kinetics process is difficult to control due to its complex and nonlinear nature and with state variables that are difficult to measure. METHODS: This paper proposes a method for automatically calculating the basal and bolus insulin doses for patients with type-1 diabetes using reinforcement learning with feedforward controller. The algorithm is designed to keep the blood glucose stable and directly compensate for the external events such as food intake. Its performance was assessed using simulation on a blood glucose model. The usage of the Kalman filter with the controller was demonstrated to estimate unmeasurable state variables. RESULTS: Comparison simulations between the proposed controller with the optimal reinforcement learning and the proportional-integral-derivative controller show that the proposed methodology has the best performance in regulating the fluctuation of the blood glucose. The proposed controller also improved the blood glucose responses and prevented hypoglycemia condition. Simulation of the control system in different uncertain conditions provided insights on how the inaccuracies of carbohydrate counting and meal-time reporting affect the performance of the control system. CONCLUSION: The proposed controller is an effective tool for reducing postmeal blood glucose rise and for countering the effects of external known events such as meal intake and maintaining blood glucose at a healthy level under uncertainties.

Comparison of computer systems and ranking criteria for automatic melanoma detection in dermoscopic images.

Melanoma is the deadliest form of skin cancer, and early detection is crucial for patient survival. Computer systems can assist in melanoma detection, but are not widespread in clinical practice. In 2016, an open challenge in classification of dermoscopic images of skin lesions was announced. A training set of 900 images with corresponding class labels and semi-automatic/manual segmentation masks was released for the challenge. An independent test set of 379 images, of which 75 were of melanomas, was used to rank the participants. This article demonstrates the impact of ranking criteria, segmentation method and classifier, and highlights the clinical perspective. We compare five different measures for diagnostic accuracy by analysing the resulting ranking of the computer systems in the challenge. Choice of performance measure had great impact on the ranking. Systems that were ranked among the top three for one measure, dropped to the bottom half when changing performance measure. Nevus Doctor, a computer system previously developed by the authors, was used to participate in the challenge, and investigate the impact of segmentation and classifier. The diagnostic accuracy when using an automatic versus the semi-automatic/manual segmentation is investigated. The unexpected small impact of segmentation method suggests that improvements of the automatic segmentation method w.r.t. resemblance to semi-automatic/manual segmentation will not improve diagnostic accuracy substantially. A small set of similar classification algorithms are used to investigate the impact of classifier on the diagnostic accuracy. The variability in diagnostic accuracy for different classifier algorithms was larger than the variability for segmentation methods, and suggests a focus for future investigations. From a clinical perspective, the misclassification of a melanoma as benign has far greater cost than the misclassification of a benign lesion. For computer systems to have clinical impact, their performance should be ranked by a high-sensitivity measure.

Using anchors from free text in electronic health records to diagnose postoperative delirium.

OBJECTIVES: Postoperative delirium is a common complication after major surgery among the elderly. Despite its potentially serious consequences, the complication often goes undetected and undiagnosed. In order to provide diagnosis support one could potentially exploit the information hidden in free text documents from electronic health records using data-driven clinical decision support tools. However, these tools depend on labeled training data and can be both time consuming and expensive to create. METHODS: The recent learning with anchors framework resolves this problem by transforming key observations (anchors) into labels. This is a promising framework, but it is heavily reliant on clinicians knowledge for specifying good anchor choices in order to perform well. In this paper we propose a novel method for specifying anchors from free text documents, following an exploratory data analysis approach based on clustering and data visualization techniques. We investigate the use of the new framework as a way to detect postoperative delirium. RESULTS: By applying the proposed method to medical data gathered from a Norwegian university hospital, we increase the area under the precision-recall curve from 0.51 to 0.96 compared to baselines. CONCLUSIONS: The proposed approach can be used as a framework for clinical decision support for postoperative delirium.

Predicting colorectal surgical complications using heterogeneous clinical data and kernel methods.

OBJECTIVE: In this work, we have developed a learning system capable of exploiting information conveyed by longitudinal Electronic Health Records (EHRs) for the prediction of a common postoperative complication, Anastomosis Leakage (AL), in a data-driven way and by fusing temporal population data from different and heterogeneous sources in the EHRs. MATERIAL AND METHODS: We used linear and non-linear kernel methods individually for each data source, and leveraging the powerful multiple kernels for their effective combination. To validate the system, we used data from the EHR of the gastrointestinal department at a university hospital. RESULTS: We first investigated the early prediction performance from each data source separately, by computing Area Under the Curve values for processed free text (0.83), blood tests (0.74), and vital signs (0.65), respectively. When exploiting the heterogeneous data sources combined using the composite kernel framework, the prediction capabilities increased considerably (0.92). Finally, posterior probabilities were evaluated for risk assessment of patients as an aid for clinicians to raise alertness at an early stage, in order to act promptly for avoiding AL complications. DISCUSSION: Machine-learning statistical model from EHR data can be useful to predict surgical complications. The combination of EHR extracted free text, blood samples values, and patient vital signs, improves the model performance. These results can be used as a framework for preoperative clinical decision support.

Support Vector Feature Selection for Early Detection of Anastomosis Leakage From Bag-of-Words in Electronic Health Records.

The free text in electronic health records (EHRs) conveys a huge amount of clinical information about health state and patient history. Despite a rapidly growing literature on the use of machine learning techniques for extracting this information, little effort has been invested toward feature selection and the features' corresponding medical interpretation. In this study, we focus on the task of early detection of anastomosis leakage (AL), a severe complication after elective surgery for colorectal cancer (CRC) surgery, using free text extracted from EHRs. We use a bag-of-words model to investigate the potential for feature selection strategies. The purpose is earlier detection of AL and prediction of AL with data generated in the EHR before the actual complication occur. Due to the high dimensionality of the data, we derive feature selection strategies using the robust support vector machine linear maximum margin classifier, by investigating: 1) a simple statistical criterion (leave-one-out-based test); 2) an intensive-computation statistical criterion (Bootstrap resampling); and 3) an advanced statistical criterion (kernel entropy). Results reveal a discriminatory power for early detection of complications after CRC (sensitivity 100%; specificity 72%). These results can be used to develop prediction models, based on EHR data, that can support surgeons and patients in the preoperative decision making phase.

Computer-Aided Decision Support for Melanoma Detection Applied on Melanocytic and Nonmelanocytic Skin Lesions: A Comparison of Two Systems Based on Automatic Analysis of Dermoscopic Images.

Commercially available clinical decision support systems (CDSSs) for skin cancer have been designed for the detection of melanoma only. Correct use of the systems requires expert knowledge, hampering their utility for nonexperts. Furthermore, there are no systems to detect other common skin cancer types, that is, nonmelanoma skin cancer (NMSC). As early diagnosis of skin cancer is essential, there is a need for a CDSS that is applicable to all types of skin lesions and is suitable for nonexperts. Nevus Doctor (ND) is a CDSS being developed by the authors. We here investigate ND's ability to detect both melanoma and NMSC and the opportunities for improvement. An independent test set of dermoscopic images of 870 skin lesions, including 44 melanomas and 101 NMSCs, were analysed by ND. Its sensitivity to melanoma and NMSC was compared to that of Mole Expert (ME), a commercially available CDSS, using the same set of lesions. ND and ME had similar sensitivity to melanoma. For ND at 95% melanoma sensitivity, the NMSC sensitivity was 100%, and the specificity was 12%. The melanomas misclassified by ND at 95% sensitivity were correctly classified by ME, and vice versa. ND is able to detect NMSC without sacrificing melanoma sensitivity.

Data-Driven Personalized Feedback to Patients with Type 1 Diabetes: A Randomized Trial.

BACKGROUND: A mobile phone-based application can be useful for patients with type 1 diabetes in managing their disease. This results in large datasets accumulated on the patient's devices, which can be used for individualized feedback. The effect of such feedback is investigated in this article. MATERIALS AND METHODS: We developed an application that included a data-driven feedback module known as Diastat for patients on self-measured blood glucose regimens. Using a stepped-wedge design, both groups initially received an application without Diastat. Group 1 activated Diastat after 4 weeks, whereas Group 2 activated Diastat 12 weeks after startup (T1). End points were glycated hemoglobin (HbA1c) level and number of out-of-range (OOR) measurements (i.e., outside the range 72-270 mg/dL). RESULTS: Thirty patients were recruited to the study, and 15 were assigned to each group after the initial meeting. There were no significant differences between groups at T1 in HbA1c or OOR events. Overall, all patients had a decrease of 0.6 percentage points in mean HbA1c (P < 0.001) and 14.5 in median OOR events over 2 weeks (P < 0.001). CONCLUSIONS: The study does not provide evidence that data-driven feedback improves glycemic control. The decrease in HbA1c was sizeable and significant, even though the study was not powered to detect this. The overall improvement in glycemic control suggests that, in general, mobile phone-based interventions can be useful in diabetes self-management.

Power law approximations of movement network data for modeling infectious disease spread.

Globalization and increased mobility of individuals enable person-to-person transmitted infectious diseases to spread faster to distant places around the world, making good models for the spread increasingly important. We study the spatiotemporal pattern of spread in the remotely located and sparsely populated region of North Norway in various models with fixed, seasonal, and random effects. The models are applied to influenza A counts using data from positive microbiology laboratory tests as proxy for the underlying disease incidence. Human travel patterns with local air, road, and sea traffic data are incorporated as well as power law approximations thereof, both with quasi-Poisson regression and based on the adjacency structure of the relevant municipalities. We investigate model extensions using information about the proportion of positive laboratory tests, data on immigration from outside North Norway and by connecting population to the movement network. Furthermore, we perform two separate analyses for nonadults and adults as children are an important driver for influenza A. Comparisons of one-step-ahead predictions generally yield better or comparable results using power law approximations.

Performance of a dermoscopy-based computer vision system for the diagnosis of pigmented skin lesions compared with visual evaluation by experienced dermatologists.

BACKGROUND: It is often difficult to differentiate early melanomas from benign melanocytic nevi even by expert dermatologists, and the task is even more challenging for primary care physicians untrained in dermatology and dermoscopy. A computer system can provide an objective and quantitative evaluation of skin lesions, reducing subjectivity in the diagnosis. OBJECTIVE: Our objective is to make a low-cost computer aided diagnostic tool applicable in primary care based on a consumer grade camera with attached dermatoscope, and compare its performance to that of experienced dermatologists. METHODS AND MATERIALS: We propose several new image-derived features computed from automatically segmented dermoscopic pictures. These are related to the asymmetry, color, border, geometry, and texture of skin lesions. The diagnostic accuracy of the system is compared with that of three dermatologists. RESULTS: With a data set of 206 skin lesions, 169 benign and 37 melanomas, the classifier was able to provide competitive sensitivity (86%) and specificity (52%) scores compared with the sensitivity (85%) and specificity (48%) of the most accurate dermatologist using only dermoscopic images. CONCLUSION: We show that simple statistical classifiers can be trained to provide a recommendation on whether a pigmented skin lesion requires biopsy to exclude skin cancer with a performance that is comparable to and exceeds that of experienced dermatologists.