2,183 research outputs found
Generating explanations from support vector machines for psychological classifications
No full textAn explanation capability is crucial in security-sensitive domains, such as medical applications. Although support vector machines (SVMs) have shown superior performance in a range of classification and regression tasks, SVMs, like artificial neural networks (ANNs), lack an explanatory capability. There is a significant literature on obtaining human-comprehensible rules from SVMs and ANNs in order to explain "how" a decision was made or "why" a certain result was achieved. This chapter proposes a novel approach to SVM classifiers. The experiments described in this chapter involve a first attempt to generate textual and visual explanations for classification results using multimedia content of various type: poems expressing positive or negative emotion, autism descriptions, and facial expressions, including those with medical relevance (facial palsy). Learned model parameters are analyzed to select important features, and filtering is applied to select feature subsets of explanatory value. The explanation components are used to generate textual summaries of classification results. We show that the explanations are consistent and that the accuracy of SVM models is bounded by the accuracy of explanation components. The results show that the generated explanations display a high level of fidelity and can generate textual summaries with an error rate of less than 35 %
Speech analysis for mental health assessment using support vector machines
No full textSpeech and language dysfunction (SLD) is one of the primary symptoms of mental disorders, such as schizophrenia. Because of the difficulties and subjective nature of SLD assessments, their use in clinical assessment of mental health problems has been limited. Recently, automated discourse analysis methods have been developed and shown the possibility of providing accurate and objective assessments more efficiently. In this chapter, we develop methods of applying Support Vector Machines (SVMs), a computational learning algorithm, in analyzing unstructured conversations of non-native English speakers, both schizophrenias and controls. In this case, the use of conventional language features, such as syntactic and semantic information, is limited because of the nature of participants: multi-cultural, non-native English speakers, and unstructured conversations. A two-level hierarchical classifier was developed that predicts specific SLD items (e.g., poverty of speech) and makes the final diagnostic decisions by combining the SLD assessment results to provide an overall assessment of the underlying mental condition. In particular, we evaluate the SVM classifiers as to their ability to predict SLD items on two mental health assessments: the Thought, Language and Communication Scale (TLC) and the Clinical Language Disorder Rating Scale (CLANG)
An alternative method of analysis in the absence of control group
No full textAlthough control groups are an important part of psychology, there are times when an appropriate control group is difficult to obtain. In the machine learning community, Support Vector Machine has often been successfully used for classification. Moreover, SVM can also be used for classification using data from one group of participants only, known as one-class SVM. In order to test the effectiveness of two-class and one-class SVMs, they were compared to TLC and CLANG in diagnosing disorganised speech. It was hypothesised that SVM would be as good as TLC and CLANG in diagnosing schizophrenic speech. It was also predicted that one-class SVM would perform just as well as two-class SVM in identifying schizophrenic speech. Lastly, it was predicted that the control group in this study would be a better match to the schizophrenic group compared to the control group studied in Chap. 11 by Tilaka. Method: 12 control group participants were each interviewed for about 20 min. The interviews were then rated for disorganised speech using TLC, CLANG, and SVM. Data for the schizophrenic group were obtained from Tilaka. Results: It was found that two-class SVM was as good as TLC and CLANG in diagnosing schizophrenic speech. It was also found that one-class SVM was comparable to two-class SVM in classifying schizophrenic speech. However, compared to the control group of Tilaka, the new control group was not a better match to the schizophrenic group. Conclusion: One-class and two-class SVMs appear to be a good alternative method of analysis
Social networks and automated mental health screening
No full textA health social network is an online information service which facilitates information sharing between closely related members of a community. The main means of finding patients with similar health conditions has been based on labor-intensive methods, such as searching the Internet. Also, because of the privacy and security issues of health information systems, it is often difficult to find patients who can support each other in the community. Over the years, many automated recommender systems have been developed for social networking. We propose a social networking framework for patient care, where health service providers facilitate social links between parents using similarities of mental health conditions. A machine learning approach was developed to automatically generate keywords for mental health descriptions that can be used to screen for mental health conditions and to then group individuals with similar mental health conditions. Keywords are generated from sources such as conversations on online forums. These keywords are then used to identify similarities between mental health descriptions, in order to recommend a community of related patients
Mental health informatics: Current approaches
No full textThe prevalence of mental disorders among both youths and adults has been growing, e.g., the number of people admitted to mental hospitals and expenditures on mental health related medical expenses have doubled over the past 10 years. Health informatics, the health applications of information technology, communication technology, and computer science generally, has long been accepted as a way to improve health services. In this paper, we review current practices and developments in mental health informatics. We broadly categorize mental health informatics into four categories: (1) telemental health (e.g., telepsychiatry), (2) automated diagnosis and assessment of mental health, (3) online mental health support (e.g., mental health social networks), and (4) mental health information management systems (e.g., electronic patient record). In particular, we review research on automated mental health assessment systems, highlighting the potential power of each in solving current mental health care problems
Automated method for diagnosing speech and language dysfunction in schizophrenia
No full textSpeech and language dysfunction (SLD) is one of the primary symptoms of schizophrenia. However, SLD measures, such as observer-rated scales, are based on clinical experience and are subjective in nature. This study compares two scales - the Thought, Language and Communication Scale (TLC) and the Clinical Language Disorder Rating Scale (CLANG) - with a novel and automated measure called Ex-Ray. The core hypothesis is that Ex-Ray either outperforms the rating scales in terms of accuracy (i.e., differentiating between schizophrenic participants and non-psychotic controls) or performs at the same level. Twenty-minute audio-recorded, unstructured interviews with 54 Singaporean participants (27 schizophrenics and 27 controls) were conducted. The interviews were rated by use of the TLC and CLANG scales. Manually transcribed texts, based on the interviews, were analysed by Ex-Ray. The three methods were then compared. Receiver Operating Characteristic (ROC) curve analysis demonstrated that Ex-Ray differentiated schizophrenic patients from normal subjects with an accuracy rate of 98 %, but did not outperform the scales at a significant level. Even though Ex-Ray is a valid and reliable measure of SLD in schizophrenia, it failed to outperform the rating scales (TLC and CLANG) for two reasons: (1) the unusually high inter-rater reliability; and (2) the uneven ethnic composition of the sample population. In a follow-up study, Ex-Ray performed at a high level even if subject and control groups were comparable in terms of [1] educational background, [2] ethnic composition (including language background) and [3] socio-economic status
Using diagnostic information to develop a machine learning application for the effective screening of autism spectrum disorders
No full textA 2-Class Support Vector Machine (SVM) classification model was developed by means of machine learning techniques and text analysis of Autism Spectrum Disorders (ASD) diagnostic reports. The ability of the 2-Class SVM application to screen for ASD is compared with other screening instruments: Gillian Autism Rating Scale—Second Edition [25], Social Communication Questionnaire [51] and Social Responsiveness Scale [11]. It was also cross-validated and refined based on a sample (n = 221). The classification performance of the SVM application was relatively better compared to the other instruments (accuracy = 83.7 %, precision = 98.8 %, sensitivity = 83.3 %, specificity = 88.9 %). A 1-Class SVM classification model was also described to highlight the usefulness of SVM with a skewed population
Suicide risk analysis
No full textThis study explores the trends and patterns in suicide risk factors using data mining techniques. Medical records of 666 suicide attempters who were admitted to a teaching hospital from January 2004 to December 2006 were studied. Data mining techniques revealed hidden patterns for repeated and single attempters, as well as suicide precipitants and risk factors. The findings have implications for further research in suicide assessment and intervention
Fin whale singalong: evidence of song conformity
No full textInternational audienceMechanisms driving song learning and conformity are still poorly known yet fundamental to understand the behavioural ecology of animals. Broadening the taxonomic range of these studies and interpreting song variation under the scope of cultural evolution will increase our knowledge on vocal learning strategies. Here, we analysed changes in fin whale (Balaenoptera physalus) songs recorded over two decades across the Central and Northeast Atlantic Ocean.We found a rapid (over 4 years) replacement of fin whale song types (different inter-note intervals -INIs) that co-existed with hybrid songs during the transition period and showed a clear geographic pattern. We also revealed gradual changes in INIs and note frequencies over more than a decade with all males adopting both rapid and gradual changes. These results provide evidence of vocal learning of rhythm in fin whale songs and conformity in both song rhythm and note frequencies.</div
Modelling the magnetic activity and filtering radial velocity curves of young Suns : the weak-line T Tauri star LkCa 4
No full textSGG acknowledges support from the Science & Technology Facilities Council (STFC) via an Ernest Rutherford Fellowship [ST/J003255/1]. SHPA acknowledges financial support from CNPq, CAPES and Fapemig.We report results of a spectropolarimetric and photometric monitoring of the weak-line T Tauri star LkCa 4 within the Magnetic Topologies of Young Stars and the Survival of close-in giant Exoplanets (MaTYSSE) programme, involving ESPaDOnS at the Canada–France–Hawaii Telescope. Despite an age of only 2 Myr and a similarity with prototypical classical T Tauri stars, LkCa 4 shows no evidence for accretion and probes an interesting transition stage for star and planet formation. Large profile distortions and Zeeman signatures are detected in the unpolarized and circularly polarized lines of LkCa 4 using Least-Squares Deconvolution (LSD), indicating the presence of brightness inhomogeneities and magnetic fields at the surface of LkCa 4. Using tomographic imaging, we reconstruct brightness and magnetic maps of LkCa 4 from sets of unpolarized and circularly polarized LSD profiles. The large-scale field is strong and mainly axisymmetric, featuring a ≃2 kG poloidal component and a ≃1 kG toroidal component encircling the star at equatorial latitudes – the latter making LkCa 4 markedly different from classical T Tauri stars of similar mass and age. The brightness map includes a dark spot overlapping the magnetic pole and a bright region at mid-latitudes – providing a good match to the contemporaneous photometry. We also find that differential rotation at the surface of LkCa 4 is small, typically ≃5.5 times weaker than that of the Sun, and compatible with solid-body rotation. Using our tomographic modelling, we are able to filter out the activity jitter in the radial velocity curve of LkCa 4 (of full amplitude 4.3 km s−1) down to an rms precision of 0.055 km s−1. Looking for hot Jupiters around young Sun-like stars thus appears feasible, even though we find no evidence for such planets around LkCa 4.Peer reviewe
- …
