Human-centred explanations for artificial intelligence systems.

As Artificial Intelligence (AI) systems increase in capability, so there are growing concerns over the ways in which the recommendations they provide can affect people's everyday life and decisions. The field of Explainable AI (XAI) aims to address such concerns but there is often a neglect of the human in this process. We present a formal definition of human-centred XAI and illustrate the application of this formalism to the design of a user interface. The user interface supports users in indicating their preferences relevant to a situation and to compare their preferences with those of a computer recommendation system. A user trial is conducted to evaluate the resulting user interface. From the user trial, we believe that users are able to appreciate how their preferences can influence computer recommendations, and how these might contrast with the preferences used by the computer. We provide guidelines of implementing human-centred XAI.

This paper presents a formal description of explanatory discourse for Human-Centred Explainable Artificial Intelligence (XAI) and demonstrate the use of this formalism as the basis for designing user interface for a recommender system. The recommender system is evaluated through a user trial. The paper concludes with guidelines for developing Human-Centred XAI.

Managing the risks associated with technological disruption in the road transport system: a control structure modelling approach.

Road transport is experiencing disruptive change from new first-of-a-kind technologies. While such technologies offer safety and operational benefits, they also pose new risks. It is critical to proactively identify risks during the design, development and testing of new technologies. The Systems Theoretic Accident Model and Processes (STAMP) method analyses the dynamic structure in place to manage safety risks. This study applied STAMP to develop a control structure model for emerging technologies in the Australian road transport system and identified control gaps. The control structure shows the actors responsible for managing risks associated with first-of-a-kind technologies and the existing control and feedback mechanisms. Gaps identified related to controls (e.g. legislation) and feedback mechanisms (e.g. monitoring for behavioural adaptation). The study provides an example of how STAMP can be used to identify control structure gaps requiring attention to support the safe introduction of new technologies.

This paper considers emerging risks associated with new technologies in the road transport system. It demonstrates a novel approach using STAMP to identify gaps in control and feedback mechanisms within the existing control structure which should be addressed to mitigate risk.

Forecasting emergent risks in advanced AI systems: an analysis of a future road transport management system.

Artificial Intelligence (AI) is being increasingly implemented within road transport systems worldwide. Next generation of AI, Artificial General Intelligence (AGI) is imminent, and is anticipated to be more powerful than current AI. AGI systems will have a broad range of abilities and be able to perform multiple cognitive tasks akin to humans that will likely produce many expected benefits, but also potential risks. This study applied the EAST Broken Links approach to forecast the functioning of an AGI system tasked with managing a road transport system and identify potential risks. In total, 363 risks were identified that could have adverse impacts on the stated goals of safety, efficiency, environmental sustainability, and economic performance of the road system. Further, risks beyond the stated goals were identified; removal from human control, mismanaging public relations, and self-preservation. A diverse set of systemic controls will be required when designing, implementing, and operating future advanced technologies.Practitioner summary: This study demonstrated the utility of HFE methods for formally considering risks associated with the design, implementation, and operation of future technologies. This study has implications for AGI research, design, and development to ensure safe and ethical AGI implementation.

Recovery characteristics of patients receiving either sugammadex or neostigmine and glycopyrrolate for reversal of neuromuscular block: a randomised controlled trial.

Sugammadex more rapidly and reliably reverses rocuronium-induced neuromuscular block compared with neostigmine, but it is not known if subsequent patient outcomes, including nausea, vomiting and other aspects of recovery are modified. In this study, we compared the recovery characteristics of sugammadex and neostigmine/glycopyrrolate following reversal of neuromuscular block. This was a single-centre, randomised, blinded, parallel-group clinical trial in women undergoing elective day-surgical laparoscopic gynaecological surgery, with a standardised general anaesthesia regimen that included rocuronium. Neuromuscular block was reversed with either sugammadex 2 mg.kg-1 or neostigmine 40 µg.kg-1 and glycopyrrolate 400 µg. The primary outcome was the incidence of nausea and vomiting during the first six postoperative hours. Secondary outcomes included other measures of postoperative recovery such as patient symptoms and recovery scores. Three-hundred and four women were analysed by intention-to-treat (sugammadex n = 151, neostigmine n = 153), which included four major protocol violations. There was no significant difference between sugammadex and neostigmine groups in the incidence of early nausea and vomiting (49.0% vs. 51.0%, respectively; OR 0.92, 95%CI 0.59-1.45; p = 0.731). Double vision (11.5% vs. 20.0%; p = 0.044) and dry mouth (71.6% vs. 85.5%; p = 0.003) were less common after sugammadex. Sedation scores at 2 h were also lower after sugammadex (median (IQR [range]) 0 (0-3 [0-10]) vs. 2 (0-4.[0-10]); p = 0.021). Twenty-four-hour recovery scores were not significantly different between groups. Reversal with sugammadex in this patient population did not reduce postoperative nausea or vomiting compared with neostigmine/glycopyrrolate.

A randomised controlled trial of parecoxib, celecoxib and paracetamol as adjuncts to patient-controlled epidural analgesia after caesarean delivery.

The benefit of combining non-opioid analgesics with neuraxial opioids for analgesia after caesarean delivery has not been clearly established. Larger doses of paracetamol or cyclooxygenase-2 inhibitors have not been evaluated. A randomised, double blind, double-dummy, parallel group placebo-controlled clinical trial was conducted among women having elective caesarean delivery under spinal anaesthesia, followed by pethidine patient-controlled epidural analgesia. Patients received placebos (group C); intravenous parecoxib 40 mg then oral celecoxib 400 mg at 12 hours (group PC); intravenous paracetamol 2 g then oral 1 g six-hourly (group PA); or these regimens combined (group PCPA). The primary outcome was 24-hour postoperative patient-controlled epidural pethidine use and the main secondary outcome was postoperative pain. One hundred and thirty-eight women were recruited but 27 subsequently met exclusion criteria, leaving 111 who were randomised, allocated and analysed by intention-to-treat (n=23, 30, 32 and 26 in groups C, PC, PA and PCPA respectively). There were no differences between groups for pethidine consumption, based on either intention-to-treat (median 365, 365, 405 and 360 mg in groups C, PC, PA and PCPA respectively, P=0.84) or per protocol analysis (17 major violations). Dynamic pain scores did not differ between groups but requirement for, and dose of, supplementary oral tramadol was least in group PCPA (incidence 23% versus 48%, 70% and 58% in groups C, PC and PA respectively, P=0.004). The addition of regular paracetamol, cyclooxygenase-2 inhibitors or both to pethidine patient-controlled epidural post-caesarean analgesia did not provide a pethidine dose-sparing effect during the first 24 hours.

Using social network analysis and agent-based modelling to explore information flow using common operational pictures for maritime search and rescue operations.

The concept of common operational pictures (COPs) is explored through the application of social network analysis (SNA) and agent-based modelling to a generic search and rescue (SAR) scenario. Comparing the command structure that might arise from standard operating procedures with the sort of structure that might arise from examining information-in-common, using SNA, shows how one structure could be more amenable to 'command' with the other being more amenable to 'control' - which is potentially more suited to complex multi-agency operations. An agent-based model is developed to examine the impact of information sharing with different forms of COPs. It is shown that networks using common relevant operational pictures (which provide subsets of relevant information to groups of agents based on shared function) could result in better sharing of information and a more resilient structure than networks that use a COP. PRACTITIONER SUMMARY: SNA and agent-based modelling are used to compare different forms of COPs for maritime SAR operations. Different forms of COP change the communications structures in the socio-technical systems in which they operate, which has implications for future design and development of a COP.

Multimodal control of sensors on multiple simulated unmanned vehicles.

The use of multimodal (speech plus manual) control of the sensors on combinations of one, two, three or five simulated unmanned vehicles (UVs) is explored. Novice controllers of simulated UVs complete a series of target checking tasks. Two experiments compare speech and gamepad control for one, two, three or five UVs in a simulated environment. Increasing the number of UVs has an impact on subjective rating of workload (measured by NASA-Task Load Index), particularly when moving from one to three UVs. Objective measures of performance showed that the participants tended to issue fewer commands as the number of vehicles increased (when using the gamepad control), but, while performance with a single UV was superior to that of multiple UVs, there was little difference across two, three or five UVs. Participants with low spatial ability (measured by the Object Perspectives Test) showed an increase in time to respond to warnings when controlling five UVs. Combining speech with gamepad control of sensors on UVs leads to superior performance on a secondary (respond-to-warnings) task (implying a reduction in demand) and use of fewer commands on primary (move-sensors and classify-target) tasks (implying more efficient operation). STATEMENT OF RELEVANCE: Benefits of multimodal control for unmanned vehicles are demonstrated. When controlling sensors on multiple UVs, participants with low spatial orientation scores have problems. It is proposed that the findings of these studies have implications for selection of UV operators and suggests that future UV workstations could benefit from multimodal control.

Measuring team skills in crime scene investigation: exploring ad hoc teams.

The purpose of this study was to measure the team skills of operational crime scene examiners (CSEs). The techniques used were based on established methods and helped to gain a greater understanding of the domain of forensic investigation. The research begins with a hierarchical task analysis and then adapts pre-established methods for measuring the performance of CSEs in four UK Police Forces. The process supports comparison between prescribed methods of 'good practice' and real world practice. This has allowed the identification of the distributed skills and tasks of the CSE. Using Annett et al.'s HTA(T), crime scene examination can be categorised and the communications and coordination structures occurring between teams investigating a burglary considered. This makes it possible to generalise the method to situations involving ad hoc teams.

Representing situation awareness in collaborative systems: a case study in the energy distribution domain.

The concept of distributed situation awareness (DSA) is currently receiving increasing attention from the human factors community. This article investigates DSA in a collaborative real-world industrial setting by discussing the results derived from a recent naturalistic study undertaken within the UK energy distribution domain. The results describe the DSA-related information used by the networks of agents involved in the scenarios analysed, the sharing of this information between the agents and the salience of different information elements used. Thus, the structure, quality and content of each network's DSA is discussed, along with the implications for DSA theory. The findings reinforce the notion that when viewing situation awareness (SA) in collaborative systems, it is useful to focus on the coordinated behaviour of the system itself, rather than on the individual as the unit of analysis and suggest that the findings from such assessments can potentially be used to inform system, procedure and training design. SA is a critical commodity for teams working in industrial systems and systems, procedures and training programmes should be designed to facilitate efficient system SA acquisition and maintenance. This article presents approaches for describing and understanding SA during real-world collaborative tasks, the outputs from which can potentially be used to inform system, training programmes and procedure design.

Distributed situation awareness in dynamic systems: theoretical development and application of an ergonomics methodology.

The purpose of this paper is to propose foundations for a theory of situation awareness based on the analysis of interactions between agents (i.e. both human and non-human) in subsystems. This approach may help to promote a better understanding of technology-mediated interaction in systems, as well as helping in the formulation of hypotheses and predictions concerning distributed situation awareness. It is proposed that agents within a system each hold their own situation awareness, which may be very different from (although compatible with) that of other agents. It is argued that we should not always hope for, or indeed want, sharing of this awareness, as different system agents have different purposes. This view marks situation awareness as a dynamic and collaborative process binding agents together on tasks on a moment-by-moment basis. Implications of this viewpoint for the development of a new theory of, and accompanying methodology for, distributed situation awareness are offered.

Human error identification techniques applied to public technology: predictions compared with observed use.

In this paper, we consider the use of human error identification (HEI) techniques as a possible alternative to observation studies for product evaluation. The HEI techniques used were Task Analysis for Error Identification (TAFEI) and Predictive Human Error Analysis (PHEA). The comparison was undertaken in connection with the prediction of errors in the use of a ticket vending machine. Two main findings emerged from the study. First, predictions derived from the HEI techniques compared favourably with errors observed in actual machine use. Second, the HEI techniques took far less time than direct observation to reach comparable levels of performance. Such rates suggest that these techniques can be usefully applied to the study of consumer products.

Arthroscopic treatment of anterolateral ankle impingement.

Fifty-five ankle arthroscopies performed for anterolateral ankle impingement between January 1987 and April 1992 were reviewed for functional outcome and patient satisfaction. There were 33 men and 22 women with an average age of 34 years (range 20-67). All patients reported a previous history of ankle injury, and 60% of the patients had at least one ankle sprain. At an average follow up of 2.6 years (range 1.0-4.5), 87% of the patients had good/excellent results, 98% were satisfied, and 84% returned to their previous sports.

An experimental comparison of test and symbols for in-car reconfigurable displays.

As cars become more sophisticated, there is a corresponding increase in the range of information which needs to be presented to drivers. This increased information load is beyond the scope of conventional technology; space requirements on the dashboard limit the number of displays, which can be employed. The study reported in this paper is based on the premise that centralized information presentation could be employed for in-car warming systems. Forty subjects were tested on a computer-based task, in which they had to rate the urgency of displayed information. The information was presented in one of four formats: symbols only, symbols plus title, symbols plus action, or symbol, title and action. Results show that the most effective type of display for urgency rating was symbol plus action, and that symbol, title and action produced the fastest reaction times. These results are considered in terms of previous research, and implications for the design of in-car warnings are discussed.

Advanced prostatic carcinoma: pulmonary manifestations.

To clarify the role of standard chest radiography in prostatic adenocarcinoma, the pulmonary manifestations of 198 patients with Stage D disease were evaluated. All patients were treated with chemotherapeutic protocols allowing for adequate clinical and radiographic correlation. Retrospective interpretation of serial chest radiographs revealed that 35% of our patients had visible intrathoracic abnormalities; however, only 24% of the patients had abnormalities attributable to intrathoracic metastases. Twenty-two percent of patients had pleural effusions, 16% reticular opacities, 3.5% reticulonodular opacities, 8% isolated or discrete pulmonary nodules, and 4.5% adenopathy. Etiologies of these opacities included metastatic disease in 93.5% of those with adenopathy and nodular or reticulonodular opacities, but 39% of pleural effusions and 52% of reticular opacities were best attributed to concomitant processes. Four patients had intrathoracic metastases without bone metastases. Standard chest radiography is a valuable screening procedure that should be correlated with clinical data to differentiate metastases from concomitant processes.

Obscured aortic arch on the lateral view as a sign of coarctation.

Forty patients with coarctation of the aorta were compared with 80 normal subjects matched by sex and age. The aortic arch was invisible on the lateral radiograph in 84% of patients preoperatively and 88% postoperatively. Only 6 normal subjects (8%) had a similar finding. The authors conclude that an obscured aortic arch on the lateral view could be a clue to the presence of coarctation. Possible explanations for this phenomenon include (a) dilatation of the brachiocephalic arteries (particularly the left subclavian artery), obscuring the superior margin of the arch; (b) hypoplasia and anteromedial displacement of the distal portion of the arch; and (c) diffuse hypoplasia of the arch combined with focal coarctation.

Protrusio acetabuli in sickle-cell anemia.

Of 155 adults with sickle-cell anemia (SS, SC), radiographs of the pelvis or hip demonstrated protrusio acetabuli on at least one side in 14 (3 men and 11 women), as indicated by projection of the acetabular line medial to the ilio-ischial line by greater than or equal to 3 mm in men and greater than or equal to 6 mm in women. All 14 patients had bone changes attributable to sickle-cell anemia, including marrow hyperplasia and osteonecrosis; however, the severity of femoral or acetabular osteonecrosis did not appear directly related to the protrusion. The authors conclude that sickle-cell anemia can predispose to development of protrusio acetabuli.

Peripheral middle lobe syndrome.

Classically, the middle lobe syndrome has been described as being caused by a central obstruction. Clinical records and radiologic findings were reviewed in 129 patients examined between 1955-1981 who had chronic disease in the right middle lobe and/or lingula. Fifty-eight patients (45%) had no evidence of a central obstructive lesion. The majority were middle-aged women with histories of chronic cough and chest pain. Bronchoscopic and radiologic evaluation served to exclude central lesions. Surgical confirmation was available in 38 patients. Pathologic study showed varying degrees of chronic inflammation, pneumonia, and bronchiectasis. Surgical results were excellent in isolated disease. Chronic atelectasis and pneumonitis of the right middle lobe and/or lingula do not always imply central obstruction. A lack of collateral ventilation is a plausible theory to explain the pathophysiology in such patients.

Differentiating empyemas and peripheral pulmonary abscesses: the value of computed tomography.

In 13 patients, conventional radiographic techniques and clinical histories were not sufficient to determine whether a peripheral cavitary lesion was an abscess or an empyema. However, after computed tomography (CT), eight patients were diagnosed as having abscesses and five as having empyemas. Abscesses had an irregular shape and a relatively thick wall which was not uniformly wide and did not have a discrete boundary between the lesion and lung parenchyma. In contrast, empyemas had a regularly shaped lumen, a smooth inner surface, and a sharply defined border between the lesion and lung. CT studies can help to distinguish between empyemas and abscesses, and treatment can be started sooner in difficult cases.