Dissertação apresentada para obtenção do grau de Mestre em Engenharia Electrotécnica e de Computadores, na Faculdade de Engenharia da Universidade do Porto

Dissertação apresentada para obtenção do grau de Mestre em Engenharia Electrotécnica e de Computadores na Faculdade de Engenharia da Universidade do Porto, sob a orientação do Prof. Doutor Adriano da Silva Carvalho

Dissertação apresentada para obtenção do grau de Mestre em Engenharia Electrotécnica e de Computadores, na Faculdade de Engenharia da Universidade do Porto, sob a orientação do Prof. Doutor José António Ruela Simões Fernandes

This was a retrospective study analyzing the diagnosis of sepsis, a severe systemic reaction to infection, in the emergency department. Sepsis is one of the leading causes of hospital mortality. Though, despite an increased focus on sepsis awareness in recent years, the rates of sepsis are increasing. Both the root causes and the bodily effects of sepsis are varied which makes screening (the identification of potentially septic patients) and diagnosis (the identification of sepsis by a medical professional) extremely difficult. In the face of this uncertainty, several attempts have been made to formalize the definition of sepsis including the systemic inflammation response syndrome (SIRS) criteria. These well-defined criteria can be used to design screens for identifying septic patients via their electronic health record (EHR), but these alerts tend to not be very selective and as such they produce many false alarms. The aim of this study was to determine how these alerts effect the decision making of physicians in the emergency department in regard sepsis diagnosis. More specifically, the goal was to determine if any of a number of well-known cognitive biases: sequential contrast effects, confirmation bias, and representativeness, could be detected in relation to sepsis diagnosis. Using a retrospective dataset of patients for which SIRS alerts were triggered, a set of behavioral criteria were designed using standard sepsis treatment procedures to determine the physicians’ diagnoses of those patients. The distribution of these diagnoses and the way past alerts were related to the diagnosis rates were analyzed. The patterns found in these analyses were constant with that would be expected in decisions made under the influence the identified biases. Additionally, there was found to be correlation between past alerts and the amount of information physicians use to make diagnoses lending further evidence of this conclusion. These results could be used to help design better alerts in the future or to improve the way medical information is presented to physicians to prevent biases from occurring in sepsis diagnosis.

The primary objective of this thesis is to study change detection problems and their applications in longitudinal and functional data. In particularly, two types of change detection problems for longitudinal data are considered. The first type of problems is on-line change detection for longitudinal data, where we focus on detecting changes within a single longitudinal data stream that arrive into the system sequentially. The other type of change detection problems that will be studied in this thesis concerns about detecting outliers from a set of longitudinal or functional data. For the first type of change detection problems, we study two novel engineering applications. The first application is studied in Chapter 2, focusing on the on-line steady state detection. The goal is to identify the transition point between transient period and steady state period. We propose a novel on-line steady state detection algorithm based on a multiple change-point state space formulation and the sequential Monte Carlo methods. Compared to other existing methods, the main contribution of this work is its significantly improved computational efficiency by the use of Rao-Blackwellization, making it a much preferred method for many on-line applications where quick processing of the data in real time is critical. Additionally, the proposed method is shown to have more robust detection performance than existing methods when dealing with different types of signals. In Chapter 3, we study the second application of change detection problems, which focuses on statistical process control for the short-run process. We propose new methods under the Bayesian framework is to track the process mean and detect on-line if the process mean is beyond certain control limits or specification limits. Our model modifies the original model proposed by Tsiamyrtzis and Hawkins (2005) and can be more flexible in handling linear trends of the process. Compared to the method proposed by Tsiamyrtzis and Hawkins (2005), the advantages of our method are two-folds. Firstly, the performance of our method is more robust to parameter misspecification and requires less knowledge of the process to make accurate estimations. Secondly, the resulted posterior inference of the process mean has a significantly reduced number of mixtures, leading to substantially save of computational and memory cost. The other type of change detection problems studied in this thesis concentrates on analysis of a set of longitudinal or functional data, which is discussed in Chapter 4. In particularly, we focus on the outlier detection for functional data, where the outlier is defined as a curve that is generated from a different process compared to normal curves. Based on the use of data depth, we propose two new depth notions, the weighted band depth and the localized weighted band depth for detecting various outliers. Our main contribution is proposing a new idea called the shape distance, which makes our methods particularly effective in detecting outliers that have different shapes with normal curves.

Twenty-two million US citizens are exposed to hazardous noise at work each year, putting them at risk for noise induced hearing loss. Noise induced hearing loss is preventable, cumulative, and irreversible with net economic impact estimated at $123 billion. While agencies such as the Occupational Safety and Health Administration have regulations in place to reduce noise induced hearing loss, these regulations are rarely enforced for agricultural workers. These workers have a low rate of hearing protection usage, with several studies finding that almost half of farmers never use hearing protection devices. Additionally, farmers have twice the hearing loss in higher frequencies and three times in mid-range frequencies than non-farmers. Use of hearing protection can reduce noise induced hearing loss, and agricultural workers are interested in increasing their usage. This makes them a promising group to target with a hearing protection intervention. This paper describes a system that combines a smartphone with a USB based noise dosimeter that can read within +/- 2 A-weighted decibels of a Class 2 sound level meter providing daily noise exposure monitoring. This device is worn by the agricultural worker throughout a work day, collecting location, accelerometer, and audio data. The data is then transferred onto the server and presented to the agricultural worker using a locally hosted website, giving personalized data of loud noise exposures that can be understood without the need for a safety specialist. The dosimeter’s data allows the agricultural worker to explore what sound pressure levels they are exposed to and get an estimate of their total noise exposure. The GPS, paired with audio clips of loud noises, allows the agricultural worker to determine what activities put them at risk of noise induced hearing loss, which are good indications of where to place hearing protection devices. The system was tested on a farm, comparing its output with several reference instruments. A-weighted, 1-second averaged sound pressure levels, GPS, and accelerometer data were collected while performing a variety of tasks indoors and outdoors. The smartphone’s external noise dosimeter read within +/- 2 dBA of the Class 2 reference dosimeter 59% of the time. The GPS devices had an average error of sub-4 meters between and the accelerometers had a mean absolute error of less than 0.1 g.

Road hazards have been threatening drivers’ safety. Drivers should perceive road hazards in order to reduce or eliminate risks (Shinar, 2007a). Hazard perception is a cognitive ability which can be improved through practice (Mark S. Horswill, Hill, & Wetton, 2015). Failures in hazard perception can cause adversities (Spainhour, Brill, John Sobanjo, Jerry Wekezer, & Primus Mtenga, 2005). Hence, understanding the cognitive process of hazard perception is important for improving drivers’ reactions and creating human-like autonomous systems. While there is no agreement among researchers about the concept of hazard perception (Baars, 1993; Mica R. Endsley, 1995b; Flach, 1994), most accepted views describe hazard perception as a process (Banbury, 2017). One of the predominant cognitive theories which describes hazard perception as a process is Neisser’s action-perception (Neisser, 1987). Neisser’s model relates hazard perception to mental map of the world and includes limitations of drivers’ memories. The action-perception model alongside other dominant hazard perception theories which describe the phenomenon can only be valid if drivers have mental models of hazards (Mica R. Endsley, 2000). Mental models which can facilitate the hazard perception process cannot be described by dominant theories that describe mental models as cognitive structural and functional models (Preece et al., 1994) In fact, the mental model which can be used for hazard perception and risk anticipation should explain how hazard can fail a traffic system and cause crash. Thus, the first experiment explores the existence of such a mental model by using Schema World Action Research Method (SWARM) cognitive probes. The result proves the existence of subjective mental models of hazards. Mental models are subjective, and drivers can have different preference of actions to a hazard accordingly. Automation of driving requires drivers to monitor the Autonomous vehicles (AVs) behaviors and takeover the control when needed (Banks & Stanton, 2017). For a successful monitoring of AVs drivers should build an appropriate level of trust in systems according to systems reliability (Lee & See, 2004). AVs should produce acceptable results to be considered reliable and drivers should develop accurate mental models of AVs actions and limitations (Walliser, 2011). Drivers will evaluate systems actions by their subjective mental models including mental models of hazards. However, AVs are designed by their designers and have limitations in replicating human-like reactions to hazards (Don Norman, 2016). The second experiment investigates how discrepancies between design models of AVs and drivers’ mental models including mental models of hazards can influence drivers’ trust in automation. In this part, naturalistic method and a Tesla S is used and participants are interviewed after five days of driving by using advanced automated systems. Results show that drivers use their mental models of hazards to predict hazardous scenarios and takeover the control before hazards materialize. Additionally, findings reveal how complexity of the system can result in function confusion, mode confusion, and misinterpreting AVs capabilities, which can result in abuse of automated systems. Results reveal that there is a need for adequate training of drivers on autonomous and advanced systems.

The increasing use of in-vehicle information systems (IVISs), such as navigation devices and MP3 players, can jeopardize safety by introducing distraction into driving. One way to address this problem is to develop distraction mitigation systems, which adapt IVIS functions according to driver state. In such a system, correctly identifying driver distraction is critical, which is the focus of this dissertation. Visual and cognitive distractions are two major types of distraction that interfere with driving most compared with other types. Visual and cognitive distraction can occur individually or in combination. The research gaps in detecting driver distraction are that the interactions of visual and cognitive distractions have not been well studied and that no accurate algorithm/strategy has been developed to detect visual, cognitive, or combined distraction. To bridge these gaps, the dissertation fulfilled three specific aims. The first aim demonstrated the layered algorithm developed based on data mining methods could improve the detection of cognitive distraction from my previous studies. The second aim developed estimation algorithms for visual distraction and demonstrated a strong relationship of the estimated distraction with the increased risk of real crashes using the naturalistic data. The third objective examined the interaction of visual and cognitive distractions and developed an effective strategy to identify combined distraction. Together these aims suggest that driver distraction can be detected from performance indicators using appropriate quantitative methods. Data mining techniques represent a promising category of methods to construct such detection algorithms. When combined in a sequential way, visual distraction dominates the effects of distraction while cognitive distraction reduces the overall impairments of distraction on driver performance. Therefore, it is not necessary to detect cognitive distraction if visual distraction is present. These approaches to detecting distraction can be also generalized to estimate other performance impairments, such as driver fatigue.

This study investigates the efficacy of Kerr’s (1957) “Goals-Freedom-Alertness” (GFA) theory of accident causation and Likert’s (1961) theory of dissimilar attitudes between supervisory levels as applied to the U.S. Forest Service. There are two types of ranger districts, those rated high and low on the basis of accident frequency ratios. From each district four subjects from each of two supervisory levels were administered a battery of ten scales. The overall results indicate that neither GFA theory was supported on the district variable, nor Likert’s theory on the supervisor variable. The results independent of theory do indicate that the manner in which the district ranger is seen by his immediate subordinates is related to accident frequency.

In this thesis, there are generally three contributions to the Ranking and Selection problem in discrete-event simulation area. Ranking and selection is an important problem when people want to select single or multiple best designs from alternative pool. There are two different types in discrete-event simulation: terminating simulation and steady-state simulation. For steady-state simulation, there is an initial trend before the data output enters into the steady-state, if we cannot start the simulation from steady state. We need to remove the initial trend before we use the data to estimate the steady-state mean. Our first contribution regards the application to eliminate the initial trend/initialization bias. In this thesis, we present a novel solution to remove the initial trend motivated by offline change detection method. The method is designed to monitor the cumulative absolute bias from the estimated steady-state mean. Experiments are conducted to compare our procedure with other existing methods. Our method is shown to be at least no worse than those methods and in some cases much better. After removing the initialization bias, we can apply a ranking and selection procedure for the data outputs from steady-state simulation. There are two main approaches to ranking and selection problem. One is subset selection and the other one is indifference zone selection. Also by employing directed graph, some single-best ranking and selection methods can be extended to solve multi-best selection problem. Our method is designed to solve multi-best ranking and selection. And in Chapter 3, one procedure for ranking and selection in terminating simulation is extended based full sequential idea. It means we compare the sample means among all systems in contention at each stage. Also, we add a technique to do pre-selection of the superior systems at the same time of eliminating inferior systems. This can accelerate the speed of obtaining the number of best systems we want. Experiments are conducted to demonstrate the pre-selection technique can save observation significantly compared with the procedure without it. Also compared with existing methods, our procedure can save significant number of observations. We also explore the effect of common random number. By using it in the simulation process, more observations can be saved. The third contribution of this thesis is to extend the procedure in Chapter 3 for steady-state simulation. Asymptotic variance is employed in this case. We justify our procedure in asymptotic point of view. And by doing extensive experiments, we demonstrate that our procedure can work in most cases when sample size is finite