Machine learning and data mining: strategies for hypothesis generation.

Strategies for generating knowledge in medicine have included observation of associations in clinical or research settings and more recently, development of pathophysiological models based on molecular biology. Although critically important, they limit hypothesis generation to an incremental pace. Machine learning and data mining are alternative approaches to identifying new vistas to pursue, as is already evident in the literature. In concert with these analytic strategies, novel approaches to data collection can enhance the hypothesis pipeline as well. In data farming, data are obtained in an 'organic' way, in the sense that it is entered by patients themselves and available for harvesting. In contrast, in evidence farming (EF), it is the provider who enters medical data about individual patients. EF differs from regular electronic medical record systems because frontline providers can use it to learn from their own past experience. In addition to the possibility of generating large databases with farming approaches, it is likely that we can further harness the power of large data sets collected using either farming or more standard techniques through implementation of data-mining and machine-learning strategies. Exploiting large databases to develop new hypotheses regarding neurobiological and genetic underpinnings of psychiatric illness is useful in itself, but also affords the opportunity to identify novel mechanisms to be targeted in drug discovery and development.

Evaluation of clinical prognostic models for suicide attempts after a major depressive episode.

OBJECTIVE: In this study, we compare the performance of prognostic models of increasing complexity for prediction of future suicide attempt. METHOD: Using data from a 2-year prospective study of 304 depressed subjects, a series of Cox proportional hazard regression models were developed to predict future suicide attempt. The models were evaluated in terms of discrimination (the ability to rank subjects in order of risk), calibration (accuracy of predicted probabilities of attempt), and sensitivity and specificity of risk group stratification based on cross-validated predicted probabilities. RESULTS: Although an additive model with past attempt, smoking status, and suicidal ideation achieved 75% (cross-validated) sensitivity and specificity, models that performed best in terms of discrimination included interactions between predictor variables. CONCLUSION: As several models had similar predictive power, clinical considerations and ease of interpretation may have a significant role in the final stage of model selection for assessing future suicide attempt risk.

Lower cortisol levels in depressed patients with comorbid post-traumatic stress disorder.

Post-traumatic stress disorder (PTSD) is often comorbid with major depressive episodes (MDEs) and both conditions carry a higher rate of suicidal behavior. Hypothalamic-pituitary-adrenal (HPA) axis and serotonin abnormalities are associated with both conditions and suicidal behavior, but their inter-relation is not known. We determined cortisol response to placebo or fenfluramine in MDE, MDE and PTSD (MDE+PTSD), and healthy volunteers (HVs) and examined the relation of cortisol responses to suicidal behavior. A total of 58 medication-free patients with MDE (13 had MDE+PTSD) and 24 HVs were studied. They received placebo on the first day and fenfluramine on the second day. Cortisol levels were drawn before challenge and for 5 h thereafter. The MDE+PTSD group had the lowest plasma cortisol, the MDE group had the highest, and HVs had intermediate levels. There were no group differences in cortisol response to fenfluramine. Suicidal behavior, sex, and childhood history of abuse were not predictors of baseline or postchallenge plasma cortisol. Cortisol levels increased with age. This study finds elevated cortisol levels in MDE and is the first report of lower cortisol levels in MDE+PTSD. The findings underscore the impact of comorbidity of PTSD with MDE and highlight the importance of considering comorbidity in psychobiology.

Experiences with ARTEMIS--an Internet-based telemedicine system.

ARTEMIS is one of the first systems to exploit the Internet/Intranet technologies for exchanging patient information among health care providers. The primary project goal was to develop and demonstrate a regional telehealth environment specifically to support real-time consultations among health care providers via a computer network, provide secure access to multi-media patient records and discharge summaries, facilitate authentication/digital sign-off, multi-media mail-based referrals, and network-based dictation/transcription. A prototype is deployed in southern West Virginia in a Community Care Network (CCN). The CCN consists of providers, hospitals, clinics, laboratories, that make up one "Virtual" clinic on the "Intranet". ARTEMIS employs new technologies such as Java and JavaScript for the browser, and CORBA-based "middleware" for interoperability at the server-end. Several experiments were designed for evaluating the impact of ARTEMIS on patient care. In this paper we discuss the challenges we faced and the means by which we plan to meet these challenges. We conclude by outlining new thrust areas in which we are concentrating in our next phase of development of ARTEMIS.

Evaluation of Community Care Network (CCN) system in a rural health care setting.

Concurrent Engineering Research Center (CERC), under the sponsorship of NLM (National Library of Medicine) is in the process of developing a computerized patient record system for a clinical environment distributed in rural West Virginia. This realization of the CCN (Community Care Network), besides providing computer-based patient records accessible from a chain of clinics and one hospital, supports collaborative health care processes like referral and consulting. To evaluate the effectiveness of the system, a study was designed and is in the process of being executed. Three surveys were designed to provide subjective measures, and four experiments for collecting objective data. Data collection is taking place in several phases: baseline data are collected before the system is deployed; the process is repeated with minimal changes three, then six months later or as often as new versions of the system are installed. Results are then to be compared, using whenever possible matching techniques (i.e. the preliminary data collected on a provider will be matched with the data collected later on the same provider). Surveys are conducted through questionnaires distributed to providers and nurses and person-to-person interviews of the patients. The time spent on patient-chart related activities is measured by work-sampling, aided by a computer application running on a laptop PC. Information about missing patient record parts is collected by the providers, the frequency by which new features of the computerized system are used will be logged by the system itself and clinical outcome measures will be studied from the results of the clinics' own patient chart audits. Preliminary results of the surveys and plans for the immediate and distant future are discussed at the end of the paper.