Peter Weller

Modeling User Preferences in Recommender Systems: A Classification Framework for Explicit and Implicit User Feedback

Recommender systems are firmly established as a standard technology for assisting users with their choices; however, little attention has been paid to the application of the user model in recommender systems, particularly the variability and noise that are an intrinsic part of human behavior and activity. To enable recommender systems to suggest items that are useful to a particular user, it can be essential to understand the user and his or her interactions with the system. These interactions typically manifest themselves as explicit and implicit user feedback that provides the key indicators for modeling users’ preferences for items and essential information for personalizing recommendations. In this article, we propose a classification framework for the use of explicit and implicit user feedback in recommender systems based on a set of distinct properties that include Cognitive Effort, User Model, Scale of Measurement, and Domain Relevance. We develop a set of comparison criteria for explicit and implicit user feedback to emphasize the key properties. Using our framework, we provide a classification of recommender systems that have addressed questions about user feedback, and we review state-of-the-art techniques to improve such user feedback and thereby improve the performance of the recommender system. Finally, we formulate challenges for future research on improvement of user feedback.

Using artificial neural networks for classifying ICU patient states

The rapid accurate diagnosis of critical disorders is an essential component of intensive care. Traditional diagnostic techniques have relied on physician experience, which is based on a data set chosen from his or her personal preferences, rather than from scientific merit. In this article, we show that there are alternative methods of selecting clinical variables on which to base a diagnosis. We suggest that a model-based technique utilizing artificial neural networks (ANNs) can be used to investigate alternative, objectively chosen data input sets. Traditionally, ANNs have been used for diagnosis or prediction tasks; however, this article introduces a novel method of exploring the inner structure of suitably trained ANNs to determine a set of key variables for each clinical state defined. Two different ANN techniques are proposed: self-organizing maps and backpropagation networks. We do not claim that these techniques provide the optimal data set for decision making, but we do show that other combinations of data exist that may be an improvement over the physician methods currently used.

Technical description of the IBIS Data Library

The IBIS Data Library (DL) is an annotated data library that contains practically all the monitored data and other clinical information from critically ill patients during surgery and in intensive care. The data have been collected at three sites: the intensive care unit of the Kuopio University Hospital, Finland; Royal Brompton Hospital, London, UK; and St. Bartholomews Hospital, London, UK. The purpose of the DL is to form the basis for development of biosignal interpretation methods in the Improved Monitoring for Brain Dysfunction in Intensive Care and Surgery project in the European Union (EU) BIOMED2 programme (BMH4-97-2570). The DL contains continuous electroencephalography signals, multimodal evoked potential recordings and diagnostic electrocardiography recorded during intensive care and surgery. In addition, signal types similar to those recorded during an earlier project, the EU-BIOMED1 project IMPROVE, are stored in the DL. In addition, trend data from patient monitors, laboratory data, annotations, nursing actions, and medications recorded and stored by a Patient Data Management System (PDMS) during routine care are included. The data obtained routinely are complemented by special annotations made by a physician who observes the patient during the data collection session. Annotations include, for example, assessment of the awareness of the patient and specific events during surgery not recorded routinely by the PDMS. Inclusion of information about the care plan and the aims of the care make the contents of the DL complete. The present paper describes the technical set-up used for recording of the DL and the contents of the DL. The paper also includes an appendix defining a new data format, the extended evoked potentials format, used for storage of sweep data in the DL.

Decision time for clinical decision support systems

Clinical decision support systems are interactive software systems designed to help clinicians with decision-making tasks, such as determining a diagnosis or recommending a treatment for a patient. Clinical decision support systems are a widely researched topic in the computer science community, but their inner workings are less well understood by, and known to, clinicians. This article provides a brief explanation of clinical decision support systems and some examples of real-world systems. It also describes some of the challenges to implementing these systems in clinical environments and posits some reasons for the limited adoption of decision-support systems in practice. It aims to engage clinicians in the development of decision support systems that can meaningfully help with their decision-making tasks and to open a discussion about the future of automated clinical decision support as a part of healthcare delivery.

Combining macula clinical signs and patient characteristics for age-related macular degeneration diagnosis: a machine learning approach

BackgroundTo investigate machine learning methods, ranging from simpler interpretable techniques to complex (non-linear) “black-box” approaches, for automated diagnosis of Age-related Macular Degeneration (AMD).MethodsData from healthy subjects and patients diagnosed with AMD or other retinal diseases were collected during routine visits via an Electronic Health Record (EHR) system. Patients’ attributes included demographics and, for each eye, presence/absence of major AMD-related clinical signs (soft drusen, retinal pigment epitelium, defects/pigment mottling, depigmentation area, subretinal haemorrhage, subretinal fluid, macula thickness, macular scar, subretinal fibrosis). Interpretable techniques known as white box methods including logistic regression and decision trees as well as less interpreitable techniques known as black box methods, such as support vector machines (SVM), random forests and AdaBoost, were used to develop models (trained and validated on unseen data) to diagnose AMD. The gold standard was confirmed diagnosis of AMD by physicians. Sensitivity, specificity and area under the receiver operating characteristic (AUC) were used to assess performance.ResultsStudy population included 487 patients (912 eyes). In terms of AUC, random forests, logistic regression and adaboost showed a mean performance of (0.92), followed by SVM and decision trees (0.90). All machine learning models identified soft drusen and age as the most discriminating variables in clinicians’ decision pathways to diagnose AMD.ConclusionsBoth black-box and white box methods performed well in identifying diagnoses of AMD and their decision pathways. Machine learning models developed through the proposed approach, relying on clinical signs identified by retinal specialists, could be embedded into EHR to provide physicians with real time (interpretable) support.

Questionnaire tools for the diagnosis of carpal tunnel syndrome from the patient history

Introduction: There remains no “gold standard” for the diagnosis of carpal tunnel syndrome (CTS). Clinical diagnosis is often held to be paramount but depends on the skills of the individual practitioner. We describe two mathematical approaches to the analysis of a history obtained by questionnaire. Methods: We used two earlier instruments, a conventional logistic regression analysis, and an artificial neural network to analyze data from 5860 patients referred for diagnosis of hand symptoms. We evaluated their ability to predict whether nerve conduction studies would show evidence of CTS using receiver operating characteristic curves. Results: Both new instruments outperformed the existing tools, achieving sensitivity of 88% and specificity of 50% in predicting abnormal median nerve conduction. When combined, 96% sensitivity and 50% specificity were achieved. Conclusion: The combined instrument can be used as a preliminary screening tool for CTS, for self‐diagnosis, and as a supplement to diagnosis in primary care. Muscle Nerve, 2011

The IBIS project: data collection in London

The primary aim of the Improved Monitoring for Brain Dysfunction during Intensive Care and Surgery (IBIS) project was to create a unique and comprehensively annotated data library (DL) of multiple physiological, including neurophysiological, signals. Data collection was undertaken in Kuopio, Finland and London, UK, and comparable protocols were used at all the sites. In London, 43 patients were recruited at the Royal Brompton Hospital, followed by nine at St. Bartholomews Hospital, all of whom underwent cardiac or combined cardiac and carotid artery surgery. Thirty-seven patients underwent a single operation, while 15 underwent two procedures. The protocols and equipment used, problems specific to the electrically hostile environment and preliminary results are described, including those of clinical interest. The DL is being used for the development of clinically applicable neurophysiological monitoring tools.

Prospective analysis of the accuracy of diagnosis of carpal tunnel syndrome using a web-based questionnaire

Objective To confirm the accuracy of a diagnostic questionnaire for carpal tunnel syndrome (CTS) when presented via a public website rather than on paper. Design Prospective comparison of the probability of CTS as assessed by the web-based questionnaire at http://www.carpal-tunnel.net with the results of nerve conduction studies. Setting Subregional neurophysiology laboratory serving a population of 700 000 in East Kent, UK. Participants 2821 individuals who were able to complete an online diagnostic questionnaire out of 4899 referred for initial diagnostic testing for new presentations with suspected CTS from April 2011 to March 2013. No exclusions were made on grounds of age, gender or coincident pathology. Main outcome measure—nerve conduction results confirming CTS. The severity of median nerve impairment demonstrated was also assessed using a validated neurophysiological scale. Results The web-based questionnaire accurately estimates the probability of CTS being confirmed on nerve conduction studies. Using a website diagnostic score of ≥40% as an example cut-off value the questionnaire achieves 78% sensitivity and 68% specificity in predicting the finding of evidence of CTS on nerve conduction studies. The web-based version of the diagnostic questionnaire was as accurate as the original paper version with an area under the receiver operating characteristic curve of 0.79. There was also a significant correlation between the diagnostic score given by the website and the severity of CTS with higher scores being associated with greater nerve dysfunction (r=0.3, p<0.00001). Conclusions Completion of the symptom questionnaire on the website by patients at home provides a prediction of the likelihood of CTS which is sufficiently accurate to be used in initial planning of investigation and treatment.

Evaluation of a wearable computer system for telemonitoring in a critical environment

This paper reports on the evaluation of a wearable computer system designed for use in a critical environment, namely the intensive care unit of a hospital. The nature of the application raised ethical issues for testing in a clinical environment and standard evaluation techniques could not easily be applied. The system was therefore evaluated by clinicians in a multi-tasking environment with a simulated set of patient scenarios. Measures of suitability and wearability were applied. The results were encouraging and the system was deemed suitable for further evaluation in the clinical setting, subject to ethical approval.

Wearable computers IN the operating room environment

High technology is a common feature in the modern operating room. While this situation enables a wide range of patient related data to be collected and analysed, the optimal viewing of this information becomes problematic. This situation is particularly acute in a busy operating theatre or while the clinician is moving around the hospital. The WINORE (Wearable computers IN the Operating Room Environment) project is a possible solution to this dilemma. It uses wearable computers and head mounted displays to provide an enhanced delivery of patient information, wirelessly collected from a range of devices, to surgeons, anaesthetists, and supervising clinicians. A crucial dimension to the project is how the clinicians interface with the system given the restrictions of sterile conditions and reduced dexterity due to operating procedures. In this paper we present the WINORE project concept, the background ideas and some findings from our trials.