Deep learning-based automatic measurement system for patellar height: a multicenter retrospective study.

BACKGROUND: The patellar height index is important; however, the measurement procedures are time-consuming and prone to significant variability among and within observers. We developed a deep learning-based automatic measurement system for the patellar height and evaluated its performance and generalization ability to accurately measure the patellar height index. METHODS: We developed a dataset containing 3,923 lateral knee X-ray images. Notably, all X-ray images were from three tertiary level A hospitals, and 2,341 cases were included in the analysis after screening. By manually labeling key points, the model was trained using the residual network (ResNet) and high-resolution network (HRNet) for human pose estimation architectures to measure the patellar height index. Various data enhancement techniques were used to enhance the robustness of the model. The root mean square error (RMSE), object keypoint similarity (OKS), and percentage of correct keypoint (PCK) metrics were used to evaluate the training results. In addition, we used the intraclass correlation coefficient (ICC) to assess the consistency between manual and automatic measurements. RESULTS: The HRNet model performed excellently in keypoint detection tasks by comparing different deep learning models. Furthermore, the pose_hrnet_w48 model was particularly outstanding in the RMSE, OKS, and PCK metrics, and the Insall-Salvati index (ISI) automatically calculated by this model was also highly consistent with the manual measurements (intraclass correlation coefficient [ICC], 0.809-0.885). This evidence demonstrates the accuracy and generalizability of this deep learning system in practical applications. CONCLUSION: We successfully developed a deep learning-based automatic measurement system for the patellar height. The system demonstrated accuracy comparable to that of experienced radiologists and a strong generalizability across different datasets. It provides an essential tool for assessing and treating knee diseases early and monitoring and rehabilitation after knee surgery. Due to the potential bias in the selection of datasets in this study, different datasets should be examined in the future to optimize the model so that it can be reliably applied in clinical practice. TRIAL REGISTRATION: The study was registered at the Medical Research Registration and Filing Information System (medicalresearch.org.cn) MR-61-23-013065. Date of registration: May 04, 2023 (retrospectively registered).

Missed connectionsThe Last Human Job: The Work of Connecting in a Disconnected World Allison J. Pugh Princeton University Press, 2024. 384 pp.

A sociologist confronts the future of care work in an age of automation.

Automatic detection of epilepsy from EEGs using a temporal convolutional network with a self-attention layer.

BACKGROUND: Over 60% of epilepsy patients globally are children, whose early diagnosis and treatment are critical for their development and can substantially reduce the disease's burden on both families and society. Numerous algorithms for automated epilepsy detection from EEGs have been proposed. Yet, the occurrence of epileptic seizures during an EEG exam cannot always be guaranteed in clinical practice. Models that exclusively use seizure EEGs for detection risk artificially enhanced performance metrics. Therefore, there is a pressing need for a universally applicable model that can perform automatic epilepsy detection in a variety of complex real-world scenarios. METHOD: To address this problem, we have devised a novel technique employing a temporal convolutional neural network with self-attention (TCN-SA). Our model comprises two primary components: a TCN for extracting time-variant features from EEG signals, followed by a self-attention (SA) layer that assigns importance to these features. By focusing on key features, our model achieves heightened classification accuracy for epilepsy detection. RESULTS: The efficacy of our model was validated on a pediatric epilepsy dataset we collected and on the Bonn dataset, attaining accuracies of 95.50% on our dataset, and 97.37% (A v. E), and 93.50% (B vs E), respectively. When compared with other deep learning architectures (temporal convolutional neural network, self-attention network, and standardized convolutional neural network) using the same datasets, our TCN-SA model demonstrated superior performance in the automated detection of epilepsy. CONCLUSION: The proven effectiveness of the TCN-SA approach substantiates its potential as a valuable tool for the automated detection of epilepsy, offering significant benefits in diverse and complex real-world clinical settings.

Automatic left ventricle volume and mass quantification from 2D cine-MRI: Investigating papillary muscle influence.

OBJECTIVE: Early detection of cardiovascular diseases is based on accurate quantification of the left ventricle (LV) function parameters. In this paper, we propose a fully automatic framework for LV volume and mass quantification from 2D-cine MR images already segmented using U-Net. METHODS: The general framework consists of three main steps: Data preparation including automatic LV localization using a convolution neural network (CNN) and application of morphological operations to exclude papillary muscles from the LV cavity. The second step consists in automatically extracting the LV contours using U-Net architecture. Finally, by integrating temporal information which is manifested by a spatial motion of myocytes as a third dimension, we calculated LV volume, LV ejection fraction (LVEF) and left ventricle mass (LVM). Based on these parameters, we detected and quantified cardiac contraction abnormalities using Python software. RESULTS: CNN was trained with 35 patients and tested on 15 patients from the ACDC database with an accuracy of 99,15 %. U-Net architecture was trained using ACDC database and evaluated using local dataset with a Dice similarity coefficient (DSC) of 99,78 % and a Hausdorff Distance (HD) of 4.468 mm (p < 0,001). Quantification results showed a strong correlation with physiological measures with a Pearson correlation coefficient (PCC) of 0,991 for LV volume, 0.962 for LVEF, 0.98 for stroke volume (SV) and 0.923 for LVM after pillars' elimination. Clinically, our method allows regional and accurate identification of pathological myocardial segments and can serve as a diagnostic aid tool of cardiac contraction abnormalities. CONCLUSION: Experimental results prove the usefulness of the proposed method for LV volume and function quantification and verify its potential clinical applicability.

A model-based cost-utility analysis of an automated notification system for deteriorating patients on general wards.

BACKGROUND: Delayed response to clinical deterioration of hospital inpatients is common. Deployment of an electronic automated advisory vital signs monitoring and notification system to signal clinical deterioration is associated with significant improvements in clinical outcomes but there is no evidence on the cost-effectiveness compared with routine monitoring, in the National Health Service (NHS) in the United Kingdom (UK). METHODS: A decision analytic model was developed to estimate the cost-effectiveness of an electronic automated advisory notification system versus standard care, in adults admitted to a district general hospital. Analyses considered: (1) the cost-effectiveness of the technology based on secondary analysis of patient level data of 3787 inpatients in a before-and-after study; and (2) the cost-utility (cost per quality-adjusted life-year (QALY)) over a lifetime horizon, extrapolated using published data. Analysis was conducted from the perspective of the NHS. Uncertainty in the model was assessed using a range of sensitivity analyses. RESULTS: The study population had a mean age of 68 years, 48% male, with a median inpatient stay of 6 days. Expected life expectancy at discharge was assumed to be 17.74 years. (1) Cost-effectiveness analysis: The automated notification system was more effective (-0.027 reduction in mean events per patient) and provided a cost saving of -£12.17 (-182.07 to 154.80) per patient admission. (2) Cost-utility analysis: Over a lifetime horizon the automated notification system was dominant, demonstrating a positive incremental QALY gain (0.0287 QALYs, equivalent to ~10 days of perfect health) and a cost saving of £55.35. At a threshold of £20,000 per QALY, the probability of automated monitoring being cost-effective in the NHS was 81%. Increased use of cableless sensors may reduce cost-savings, however, the intervention remains cost-effective at 100% usage (ICER: £3,107/QALY). Stratified cost-effectiveness analysis by age, National Early Warning Score (NEWS) on admission, and primary diagnosis indicated the automated notification system was cost-effective for most strategies and that use representative of the patient population studied was the most cost-saving strategy. CONCLUSION: Automated notification system for adult patients admitted to general wards appears to be a cost-effective use in the NHS; adopting this technology could be good use of scarce resources with significance for patient safety.

The fine line between automation and augmentation in website usability evaluation.

Artificial Intelligence (AI) systems are becoming widespread in all aspects of society, bringing benefits to the whole economy. There is a growing understanding of the potential benefits and risks of this type of technology. While the benefits are more efficient decision processes and industrial productivity, the risks may include a potential progressive disengagement of human beings in crucial aspects of decision-making. In this respect, a new perspective is emerging that aims at reconsidering the centrality of human beings while reaping the benefits of AI systems to augment rather than replace professional skills: Human-Centred AI (HCAI) is a novel framework that posits that high levels of human control do not contradict high levels of computer automation. In this paper, we investigate the two antipodes, automation vs augmentation, in the context of website usability evaluation. Specifically, we have analyzed whether the level of automation provided by a tool for semi-automatic usability evaluation can support evaluators in identifying usability problems. Three different visualizations, each one corresponding to a different level of automation, ranging from a full-automation approach to an augmentation approach, were compared in an experimental study. We found that a fully automated approach could help evaluators detect a significant number of medium and high-severity usability problems, which are the most critical in a software system; however, it also emerged that it was possible to detect more low-severity usability problems using one of the augmented approaches proposed in this paper.

Delivering on a promise: futureproofing automated insect monitoring methods.

Due to rapid technological innovations, the automated monitoring of insect assemblages comes within reach. However, this continuous innovation endangers the methodological continuity needed for calculating reliable biodiversity trends in the future. Maintaining methodological continuity over prolonged periods of time is not trivial, since technology improves, reference libraries grow and both the hard- and software used now may no longer be available in the future. Moreover, because data on many species are collected at the same time, there will be no simple way of calibrating the outputs of old and new devices. To ensure that reliable long-term biodiversity trends can be calculated using the collected data, I make four recommendations: (1) Construct devices to last for decades, and have a five-year overlap period when devices are replaced. (2) Construct new devices to resemble the old ones, especially when some kind of attractant (e.g. light) is used. Keep extremely detailed metadata on collection, detection and identification methods, including attractants, to enable this. (3) Store the raw data (sounds, images, DNA extracts, radar/lidar detections) for future reprocessing with updated classification systems. (4) Enable forward and backward compatibility of the processed data, for example by in-silico data 'degradation' to match the older data quality. This article is part of the theme issue 'Towards a toolkit for global insect biodiversity monitoring'.

A qualitative interview study to determine barriers and facilitators of implementing automated decision support tools for genomic data access.

Data access committees (DAC) gatekeep access to secured genomic and related health datasets yet are challenged to keep pace with the rising volume and complexity of data generation. Automated decision support (ADS) systems have been shown to support consistency, compliance, and coordination of data access review decisions. However, we lack understanding of how DAC members perceive the value add of ADS, if any, on the quality and effectiveness of their reviews. In this qualitative study, we report findings from 13 semi-structured interviews with DAC members from around the world to identify relevant barriers and facilitators to implementing ADS for genomic data access management. Participants generally supported pilot studies that test ADS performance, for example in cataloging data types, verifying user credentials and tagging datasets for use terms. Concerns related to over-automation, lack of human oversight, low prioritization, and misalignment with institutional missions tempered enthusiasm for ADS among the DAC members we engaged. Tensions for change in institutional settings within which DACs operated was a powerful motivator for why DAC members considered the implementation of ADS into their access workflows, as well as perceptions of the relative advantage of ADS over the status quo. Future research is needed to build the evidence base around the comparative effectiveness and decisional outcomes of institutions that do/not use ADS into their workflows.

Automatic offline-capable smartphone paper-based microfluidic device for efficient biomarker detection of Alzheimer's disease.

BACKGROUND: Alzheimer's disease (AD) is a prevalent neurodegenerative disease with no effective treatment. Efficient and rapid detection plays a crucial role in mitigating and managing AD progression. Deep learning-assisted smartphone-based microfluidic paper analysis devices (µPADs) offer the advantages of low cost, good sensitivity, and rapid detection, providing a strategic pathway to address large-scale disease screening in resource-limited areas. However, existing smartphone-based detection platforms usually rely on large devices or cloud servers for data transfer and processing. Additionally, the implementation of automated colorimetric enzyme-linked immunoassay (c-ELISA) on µPADs can further facilitate the realization of smartphone µPADs platforms for efficient disease detection. RESULTS: This paper introduces a new deep learning-assisted offline smartphone platform for early AD screening, offering rapid disease detection in low-resource areas. The proposed platform features a simple mechanical rotating structure controlled by a smartphone, enabling fully automated c-ELISA on µPADs. Our platform successfully applied sandwich c-ELISA for detecting the ß-amyloid peptide 1-42 (Aß 1-42, a crucial AD biomarker) and demonstrated its efficacy in 38 artificial plasma samples (healthy: 19, unhealthy: 19, N = 6). Moreover, we employed the YOLOv5 deep learning model and achieved an impressive 97 % accuracy on a dataset of 1824 images, which is 10.16 % higher than the traditional method of curve-fitting results. The trained YOLOv5 model was seamlessly integrated into the smartphone using the NCNN (Tencent's Neural Network Inference Framework), enabling deep learning-assisted offline detection. A user-friendly smartphone application was developed to control the entire process, realizing a streamlined "samples in, answers out" approach. SIGNIFICANCE: This deep learning-assisted, low-cost, user-friendly, highly stable, and rapid-response automated offline smartphone-based detection platform represents a good advancement in point-of-care testing (POCT). Moreover, our platform provides a feasible approach for efficient AD detection by examining the level of Aß 1-42, particularly in areas with low resources and limited communication infrastructure.

Semi-automated protocol to quantify and characterize fluorescent three-dimensional vascular images.

The microvasculature facilitates gas exchange, provides nutrients to cells, and regulates blood flow in response to stimuli. Vascular abnormalities are an indicator of pathology for various conditions, such as compromised vessel integrity in small vessel disease and angiogenesis in tumors. Traditional immunohistochemistry enables the visualization of tissue cross-sections containing exogenously labeled vasculature. Although this approach can be utilized to quantify vascular changes within small fields of view, it is not a practical way to study the vasculature on the scale of whole organs. Three-dimensional (3D) imaging presents a more appropriate method to visualize the vascular architecture in tissue. Here we describe the complete protocol that we use to characterize the vasculature of different organs in mice encompassing the methods to fluorescently label vessels, optically clear tissue, collect 3D vascular images, and quantify these vascular images with a semi-automated approach. To validate the automated segmentation of vascular images, one user manually segmented one hundred random regions of interest across different vascular images. The automated segmentation results had an average sensitivity of 83±11% and an average specificity of 91±6% when compared to manual segmentation. Applying this procedure of image analysis presents a method to reliably quantify and characterize vascular networks in a timely fashion. This procedure is also applicable to other methods of tissue clearing and vascular labels that generate 3D images of microvasculature.

sEMG-based automatic characterization of swallowed materials.

Monitoring of ingestive activities is critically important for managing the health and wellness of individuals with various health conditions, including the elderly, diabetics, and individuals seeking better weight control. Monitoring swallowing events can be an ideal surrogate for developing streamlined methods for effective monitoring and quantification of eating or drinking events. Swallowing is an essential process for maintaining life. This seemingly simple process is the result of coordinated actions of several muscles and nerves in a complex fashion. In this study, we introduce automated methods for the detection and quantification of various eating and drinking activities. Wireless surface electromyography (sEMG) was used to detect chewing and swallowing from sEMG signals obtained from the sternocleidomastoid muscle, in addition to signals obtained from a wrist-mounted IMU sensor. A total of 4675 swallows were collected from 55 participants in the study. Multiple methods were employed to estimate bolus volumes in the case of fluid intake, including regression and classification models. Among the tested models, neural networks-based regression achieved an R2 of 0.88 and a root mean squared error of 0.2 (minimum bolus volume was 10 ml). Convolutional neural networks-based classification (when considering each bolus volume as a separate class) achieved an accuracy of over 99% using random cross-validation and around 66% using cross-subject validation. Multiple classification methods were also used for solid bolus type detection, including SVM and decision trees (DT), which achieved an accuracy above 99% with random validation and above 94% in cross-subject validation. Finally, regression models with both random and cross-subject validation were used for estimating the solid bolus volume with an R2 value that approached 1 and root mean squared error values as low as 0.00037 (minimum solid bolus weight was 3 gm). These reported results lay the foundation for a cost-effective and non-invasive method for monitoring swallowing activities which can be extremely beneficial in managing various chronic health conditions, such as diabetes and obesity.

Automated analyses of dried blood spots collected by volumetric microsampling devices.

BACKGROUND: Dried blood spot (DBS) sampling on cellulose cards suffers from varying blood haematocrit levels and from chromatographic effects, which have a direct impact on quantitative DBS analyses. Commercial volumetric microsampling devices were, therefore, introduced to mitigate these effects, however, these devices are not compatible with automated DBS processing systems and must be processed manually. RESULTS: Capillary electrophoresis (CE) instruments use fused-silica (FS) capillaries for precise and accurate liquid handling as well as for injection, separation, and quantitative analyses of liquid samples. These inherent features of an Agilent 7100 CE instrument were employed for the automated processing (elution and homogenization) of DBSs collected by hemaPEN® volumetric devices (2.74 µL of capillary blood per spot). The hemaPEN® samples were processed directly in CE vials by consecutive transfers of 56 µL of methanol and 14 µL of deionized water through the FS capillary in a sequence of 39 DBSs with repeatability of the liquid transfers better than 1.4 %. The resulting DBS eluates were homogenized by a quick air flush through the capillary and analyzed by the same capillary and CE instrument. Creatinine was selected as a clinically relevant model analyte and its endogenous concentrations in DBSs were determined by CE with capacitively coupled contactless conductivity detection (CE-C4D) in a background electrolyte solution consisting of 50 mM acetic acid and 0.1 % (v/v) Tween 20 (pH 3.0). The overall repeatability of the automated DBS processing and CE-C4D analyses of 39 DBSs was ≤7.1 % (peak areas) and ≤0.6 % (migration times), the calibration curve was linear in the 25-500 µM range (R2 = 0.9993) and covered all endogenous blood creatinine levels, the limit of detection was 5.0 µM, and sample throughput was >12 DBSs per hour. DBS ageing for 60 days and varying blood haematocrit levels (20-70 %) did not affect creatinine quantitative results (≤6.9 % for peak areas). Inter-capillary and inter-instrument repeatability was ≤7.7 % (peak areas) and ≤3.4 % (migration times) and demonstrated an excellent transferability of the proposed analytical concept among laboratories. SIGNIFICANCE AND NOVELTY: This contribution is the first-ever report on the use of a single off-the-shelf analytical instrument for fully automated analyses of DBSs collected by commercial volumetric microsampling devices and holds great promise for future unmanned quantitative DBS analyses.

An fNIRS dataset for driving risk cognition of passengers in highly automated driving scenarios.

For highly autonomous vehicles, human does not need to operate continuously vehicles. The brain-computer interface system in autonomous vehicles will highly depend on the brain states of passengers rather than those of human drivers. It is a meaningful and vital choice to translate the mental activities of human beings, essentially playing the role of advanced sensors, into safe driving. Quantifying the driving risk cognition of passengers is a basic step toward this end. This study reports the creation of an fNIRS dataset focusing on the prefrontal cortex activity in fourteen types of highly automated driving scenarios. This dataset considers age, sex and driving experience factors and contains the data collected from an 8-channel fNIRS device and the data of driving scenarios. The dataset provides data support for distinguishing the driving risk in highly automated driving scenarios via brain-computer interface systems, and it also provides the possibility of preventing potential hazards in some scenarios, in which risk remains at a high value for an extended period, before hazard occurs.

A new chromatographic response function with automatically adapting weight factor for automated method development.

A critical factor for automated method development in chromatography is the maximization or minimization of an objective function describing the quality (and speed) of the separation. In chromatography, this function is commonly referred to as a chromatographic response function (CRF). Many CRFs have previously been introduced, but many have unfavourable properties such as featuring multiple optima, insufficient discriminatory power, and a too strong dependence on the weight factors needed to balance resolution and time penalty components. To overcome these problems, the present study introduces a new type of CRF wherein the relative weight of the time penalty term is a self-adaptive function of the separation quality. The ability to unambiguously identify the optimal gradient settings of this newly proposed CRF is compared to that of some of the most frequently used CRFs in a study covering 100 randomly composed in silico samples. Doing so, the new CRF is found to flawlessly lead to the correct solution (=linear gradient parameters providing the highest resolution in the shortest potential time) in 100 % of the cases, while the most frequently used literature CRFs were off-target for about 50 to 60 % of the samples, even when considering the availability of spectral peak shape data. Some slight alterations to the proposed CRF are introduced and discussed as well.

Finely-Tuning Chemiluminescent Color of CdTe Nanocrystals and Its Application for Near-Infrared Semi-Automatic Immunoassay.

Chemiluminescence (CL), especially commercialized CL immunoassay (CLIA), is normally performed within the eye-visible region of the spectrum by exploiting the electronic-transition-related emission of the molecule luminophore. Herein, dual-stabilizers-capped CdTe nanocrystals (NCs) is employed as a model of nanoparticulated luminophore to finely tune the CL color with superior color purity. Initialized by oxidizing the CdTe NCs with potassium periodate (KIO4), intermediates of the reactive oxygen species (ROS) tend to charge CdTe NCs in both series-connection and parallel-connection routes and dominate the charge-transfer CL of CdTe NCs. The CdTe NCs/KIO4 system can exhibit color-tunable CL with the maximum emission wavelength shifted from 694 nm to 801 nm, and the red-shift span is over 100 nm. Both PL and CL of each of the CdTe NCs are bandgap-engineered; the change in the NCs surface state via CL reaction enables CL of each of the CdTe NCs to be red-shifted for ∼20 nm to PL, while the change in the NCs surface state via labeling CdTe NCs to secondary-antibody (Ab2) enables CL of the CdTe NCs-Ab2 conjugates to be red-shifted for another ∼20 nm to bare CdTe NCs. The CL of CdTe753-Ab2/KIO4 is ∼791 nm, which can perform near-infrared CL immunoassay and semi-automatically determined procalcitonin (PCT) on commercialized in vitro diagnosis (IVD) instruments.

Automated escape system: identifying prey's kinematic and behavioral features critical for predator evasion.

Identifying the kinematic and behavioral variables of prey that influence evasion from predator attacks remains challenging. To address this challenge, we have developed an automated escape system that responds quickly to an approaching predator and pulls the prey away from the predator rapidly, similar to real prey. Reaction distance, response latency, escape speed and other variables can be adjusted in the system. By repeatedly measuring the response latency and escape speed of the system, we demonstrated the system's ability to exhibit fast and rapid responses while maintaining consistency across successive trials. Using the live predatory fish species Coreoperca kawamebari, we show that escape speed and reaction distance significantly affect the outcome of predator-prey interactions. These findings indicate that the developed escape system is useful for identifying kinematic and behavioral features of prey that are critical for predator evasion, as well as for measuring the performance of predators.

Automatically detecting OSAHS patients based on transfer learning and model fusion.

Objective. Snoring is the most typical symptom of obstructive sleep apnea hypopnea syndrome (OSAHS) that can be used to develop a non-invasive approach for automatically detecting OSAHS patients.Approach. In this work, a model based on transfer learning and model fusion was applied to classify simple snorers and OSAHS patients. Three kinds of basic models were constructed based on pretrained Visual Geometry Group-16 (VGG16), pretrained audio neural networks (PANN), and Mel-frequency cepstral coefficient (MFCC). The XGBoost was used to select features based on feature importance, the majority voting strategy was applied to fuse these basic models and leave-one-subject-out cross validation was used to evaluate the proposed model.Main results. The results show that the fused model embedded with top-5 VGG16 features, top-5 PANN features, and MFCC feature can correctly identify OSAHS patients (AHI > 5) with 100% accuracy.Significance. The proposed fused model provides a good classification performance with lower computational cost and higher robustness that makes detecting OSAHS patients at home possible.

Life cycle assessment and cost analysis of an innovative automatic system for sorting municipal solid waste: A case study at Milan Malpensa airport.

With the recent advancement in artificial intelligence, there are new opportunities to adopt smart technologies for the sorting of materials at the beginning of the recycling value chain. An automatic bin capable of sorting the waste among paper, plastic, glass & aluminium, and residual waste was installed in public areas of Milan Malpensa airport, a context where the separate collection is challenging. First, the airport waste composition was assessed, together with the efficiency of the manual sorting performed by passengers among the conventional bins: paper, plastic, glass & aluminium, and residual waste. Then, the environmental (via the life cycle assessment - LCA) and the economic performances of the current system were compared to those of a system in which the sorting is performed by the automatic bin. Three scenarios were evaluated: i) all waste from public areas, despite being separately collected, is sent to incineration with energy recovery, due to the inadequate separation quality (S0); ii) recyclable fractions are sent to recycling according to the actual level of impurities in the bags (S0R); iii) fractions are sorted by the automatic bin and sent to recycling (S1). According to the results, the current separate collection shows a 62 % classification accuracy. Focusing on LCA, S0 causes an additional burden of 12.4 mPt (milli points) per tonne of waste. By contrast, S0R shows a benefit (-26.4 mPt/t) and S1 allows for a further 33 % increase of benefits. Moreover, the cost analysis indicates potential savings of 24.3 €/t in S1, when compared to S0.

Predicting driver's takeover time based on individual characteristics, external environment, and situation awareness.

The driver's takeover time is crucial to ensure a safe takeover transition in conditional automated driving. The study aimed to construct a prediction model of driver's takeover time based on individual characteristics, external environment, and situation awareness variables. A total of 18 takeover events were designed with scenarios, non-driving-related tasks, takeover request time, and traffic flow as variables. High-fidelity driving simulation experiments were carried out, through which the driver's takeover data was obtained. Fifteen basic factors and three dynamic factors were extracted from individual characteristics, external environment, and situation awareness. In this experiment, these 18 factors were selected as input variables, and XGBoost and Shapely were used as prediction methods. A takeover time prediction model (BM + SA model) was then constructed. Moreover, we analyzed the main effect of input variables on takeover time, and the interactive contribution made by the variables. And in this experiment, the 15 basic factors were selected as input variables, and the basic takeover time prediction model (BM model) was constructed. In addition, this study compared the performance of the two models and analyzed the contribution of input variables to takeover time. The results showed that the goodness of fit of the BM + SA model (Adjusted_R2) was 0.7746. The XGBoost model performs better than other models (support vector machine, random forest, CatBoost, and LightBoost models). The relative importance degree of situation awareness variables, individual characteristic variables, and external environment variables to takeover time gradually reduced. Takeover time increased with the scan and gaze durations and decreased with pupil area and self-reported situation awareness scores. There was also an interaction effect between the variables to affect takeover time. Overall, the performance of the BM + SA model was better than that of the BM model. This study can provide support for predicting driver's takeover time and analyzing the mechanism of influence on takeover time. This study can provide support for the development of real-time driver's takeover ability prediction systems and optimization of human-machine interaction design in automated vehicles, as well as for the management department to evaluate and improve the driver's takeover performance in a targeted manner.

Safety in higher level automated vehicles: Investigating edge cases in crashes of vehicles equipped with automated driving systems.

With emerging Automated Driving Systems (ADS) representing Automated Vehicles (AVs) of Level 3 or higher as classified by the Society of Automotive Engineers, several AV manufacturers are testing their vehicles on public roadways in the U.S. The safety performance of AVs has become a major concern for the transportation industry. Several ADS-equipped vehicle crashes have been reported to the National Highway Traffic Safety Administration (NHTSA) in recent years. Scrutinizing these crashes can reveal rare or complex scenarios beyond the normal capabilities of AV technologies called "edge cases." Investigating edge-case crashes helps AV companies prepare vehicles to handle these unusual scenarios and, as such, improves traffic safety. Through analyzing the NHTSA data from July 2021 to February 2023, this study utilizes an unsupervised machine learning technique, hierarchical clustering, to identify edge cases in ADS-equipped vehicle crashes. Fifteen out of 189 observations are identified as edge cases, representing 8 % of the population. Injuries occurred in 10 % of all crashes (19 out of 189), but the proportion rose to 27 % for edge cases (4 out of 15 edge cases). Based on the results, edge cases could be initiated by AVs, humans, infrastructure/environment, or their combination. Humans can be identified as one of the contributors to the onset of edge-case crashes in 60 % of the edge cases (9 out of 15 edge cases). The main scenarios for edge cases include unlawful behaviors of crash partners, absence of a safety driver within the AV, precrash disengagement, and complex events challenging for ADS, e.g., unexpected obstacles, unclear road markings, and sudden and unexpected changes in traffic flow, such as abrupt road congestion or sudden stopped traffic from a crash. Identifying and investigating edge cases is crucial for improving transportation safety and building public trust in AVs.