Stabilized sparse ordinal regression for medical risk stratification

The recent wide adoption of electronic medical records (EMRs) presents great opportunities and challenges for data mining. The EMR data are largely temporal, often noisy, irregular and high dimensional. This paper constructs a novel ordinal regression framework for predicting medical risk stratification from EMR. First, a conceptual view of EMR as a temporal image is constructed to extract a diverse set of features. Second, ordinal modeling is applied for predicting cumulative or progressive risk. The challenges are building a transparent predictive model that works with a large number of weakly predictive features, and at the same time, is stable against resampling variations. Our solution employs sparsity methods that are stabilized through domain-specific feature interaction networks. We introduces two indices that measure the model stability against data resampling. Feature networks are used to generate two multivariate Gaussian priors with sparse precision matrices (the Laplacian and Random Walk). We apply the framework on a large short-term suicide risk prediction problem and demonstrate that our methods outperform clinicians to a large margin, discover suicide risk factors that conform with mental health knowledge, and produce models with enhanced stability. © 2014 Springer-Verlag London.

FRAX (Aus) and falls risk: Association in men and women

Purpose: The WHO fracture risk prediction tool (FRAX®) utilises clinical risk factors to estimate the probability of fracture over a 10-year period. Although falls increase fracture risk, they have not been incorporated into FRAX. It is currently unclear if FRAX captures falls risk and whether addition of falls would improve fracture prediction. We aimed to investigate the association of falls risk and Australian-specific FRAX. Methods: Clinical risk factors were documented for 735 men and 602 women (age 40-90. yr) assessed at follow-up (2006-2010 and 2000-2003, respectively) of the Geelong Osteoporosis Study. FRAX scores with and without BMD were calculated. A falls risk score was determined at the time of BMD assessment and self-reported incident falls were documented from questionnaires returned one year later. Multivariable analyses were performed to determine: (i) cross-sectional association between FRAX scores and falls risk score (Elderly Falls Screening Test, EFST) and (ii) prospective relationship between FRAX and time to a fall. Results: There was an association between FRAX (hip with BMD) and EFST scores (. β=. 0.07, p<. 0.001). After adjustment for sex and age, the relationship became non-significant (. β=. 0.00, p=. 0.79). The risk of incident falls increased with increasing FRAX (hip with BMD) score (unadjusted HR 1.04, 95% CI 1.02, 1.07). After adjustment for age and sex, the relationship became non-significant (1.01, 95% CI 0.97, 1.05). Conclusions: There is a weak positive correlation between FRAX and falls risk score, that is likely explained by the inclusion of age and sex in the FRAX model. These data suggest that FRAX score may not be a robust surrogate for falls risk and that inclusion of falls in fracture risk assessment should be further explored.

Improved risk predictions via sparse imputation of patient conditions in electronic medical records

Electronic Medical Records (EMR) are increasingly used for risk prediction. EMR analysis is complicated by missing entries. There are two reasons - the “primary reason for admission” is included in EMR, but the co-morbidities (other chronic diseases) are left uncoded, and, many zero values in the data are accurate, reflecting that a patient has not accessed medical facilities. A key challenge is to deal with the peculiarities of this data - unlike many other datasets, EMR is sparse, reflecting the fact that patients have some, but not all diseases. We propose a novel model to fill-in these missing values, and use the new representation for prediction of key hospital events. To “fill-in” missing values, we represent the feature-patient matrix as a product of two low rank factors, preserving the sparsity property in the product. Intuitively, the product regularization allows sparse imputation of patient conditions reflecting common comorbidities across patients. We develop a scalable optimization algorithm based on Block coordinate descent method to find an optimal solution. We evaluate the proposed framework on two real world EMR cohorts: Cancer (7000 admissions) and Acute Myocardial Infarction (2652 admissions). Our result shows that the AUC for 3 months admission prediction is improved significantly from (0.741 to 0.786) for Cancer data and (0.678 to 0.724) for AMI data. We also extend the proposed method to a supervised model for predicting of multiple related risk outcomes (e.g. emergency presentations and admissions in hospital over 3, 6 and 12 months period) in an integrated framework. For this model, the AUC averaged over outcomes is improved significantly from (0.768 to 0.806) for Cancer data and (0.685 to 0.748) for AMI data.

Is Everyone Rated Equal? An Examination of Factors Related to Sexual Risk in Ethnically Diverse Male Adolescents Who Have Sexually Offended

As the juvenile justice system has evolved, there has been a need for clinicians to make judgments about risk posed by adolescents who have committed sexual offenses. There are inherent difficulties in attempting to assess risk for violence among adolescents due to the developmental changes taking place and the absence of well-validated instruments to guide risk prediction judgments. With minority groups increasing in numbers in the U.S., it is likely that professionals will encounter minority individuals when conducting risk assessments. Overall questions regarding race/ethnicity have been neglected and there are few if any published research that explores risk factors with minority juvenile sex offenders. The present study examined whether differences exist between Caucasian and racial/ethnic minority adolescent sexual offenders on four risk assessment measures (J-SORRAT-II, J-SOAP-II, SAVRY, and ERASOR). The sample of 207 male adolescent sexual offenders was drawn from treatment facilities in a Midwestern state. Overall results indicated that minority adolescent sex offenders had fewer risk factors endorsed than Caucasian youth across all risk assessment tools. Exploration of interactions between race and factors such as: family status, exposure to family violence, and family history of criminality upon the assessment tools risk ratings yielded non-significant findings. Limitations, suggestions for future directions, and clinical implications are discussed.

Prediction of first coronary events with the Framingham score: a systematic review

BACKGROUND: Uncertainty exists about the performance of the Framingham risk score when applied in different populations. OBJECTIVE: We assessed calibration of the Framingham risk score (ie, relationship between predicted and observed coronary event rates) in US and non-US populations free of cardiovascular disease. METHODS: We reviewed studies that evaluated the performance of the Framingham risk score to predict first coronary events in a validation cohort, as identified by Medline, EMBASE, BIOSIS, and Cochrane library searches (through August 2005). Two reviewers independently assessed 1496 studies for eligibility, extracted data, and performed quality assessment using predefined forms. RESULTS: We included 25 validation cohorts of different population groups (n = 128,000) in our main analysis. Calibration varied over a wide range from under- to overprediction of absolute risk by factors of 0.57 to 2.7. Risk prediction for 7 cohorts (n = 18658) from the United States, Australia, and New Zealand was well calibrated (corresponding figures: 0.87-1.08; for the 5 biggest cohorts). The estimated population risks for first coronary events were strongly associated (goodness of fit: R2 = 0.84) and in good agreement with observed risks (coefficient for predicted risk: beta = 0.84; 95% CI 0.41-1.26). In 18 European cohorts (n = 109499), the corresponding figures indicated close association (R2 = 0.72) but substantial overprediction (beta = 0.58, 95% CI 0.39-0.77). The risk score was well calibrated on the intercept for both population clusters. CONCLUSION: The Framingham score is well calibrated to predict first coronary events in populations from the United States, Australia, and New Zealand. Overestimation of absolute risk in European cohorts requires recalibration procedures.

Risk of severe asthma episodes predicted from fluctuation analysis of airway function

Asthma is an increasing health problem worldwide, but the long-term temporal pattern of clinical symptoms is not understood and predicting asthma episodes is not generally possible. We analyse the time series of peak expiratory flows, a standard measurement of airway function that has been assessed twice daily in a large asthmatic population during a long-term crossover clinical trial. Here we introduce an approach to predict the risk of worsening airflow obstruction by calculating the conditional probability that, given the current airway condition, a severe obstruction will occur within 30 days. We find that, compared with a placebo, a regular long-acting bronchodilator (salmeterol) that is widely used to improve asthma control decreases the risk of airway obstruction. Unexpectedly, however, a regular short-acting beta2-agonist bronchodilator (albuterol) increases this risk. Furthermore, we find that the time series of peak expiratory flows show long-range correlations that change significantly with disease severity, approaching a random process with increased variability in the most severe cases. Using a nonlinear stochastic model, we show that both the increased variability and the loss of correlations augment the risk of unstable airway function. The characterization of fluctuations in airway function provides a quantitative basis for objective risk prediction of asthma episodes and for evaluating the effectiveness of therapy.

Evolving biomarkers improve prediction of long-term mortality in patients with stable coronary artery disease: the BIO-VILCAD score.

OBJECTIVE Algorithms to predict the future long-term risk of patients with stable coronary artery disease (CAD) are rare. The VIenna and Ludwigshafen CAD (VILCAD) risk score was one of the first scores specifically tailored for this clinically important patient population. The aim of this study was to refine risk prediction in stable CAD creating a new prediction model encompassing various pathophysiological pathways. Therefore, we assessed the predictive power of 135 novel biomarkers for long-term mortality in patients with stable CAD. DESIGN, SETTING AND SUBJECTS We included 1275 patients with stable CAD from the LUdwigshafen RIsk and Cardiovascular health study with a median follow-up of 9.8 years to investigate whether the predictive power of the VILCAD score could be improved by the addition of novel biomarkers. Additional biomarkers were selected in a bootstrapping procedure based on Cox regression to determine the most informative predictors of mortality. RESULTS The final multivariable model encompassed nine clinical and biochemical markers: age, sex, left ventricular ejection fraction (LVEF), heart rate, N-terminal pro-brain natriuretic peptide, cystatin C, renin, 25OH-vitamin D3 and haemoglobin A1c. The extended VILCAD biomarker score achieved a significantly improved C-statistic (0.78 vs. 0.73; P = 0.035) and net reclassification index (14.9%; P < 0.001) compared to the original VILCAD score. Omitting LVEF, which might not be readily measureable in clinical practice, slightly reduced the accuracy of the new BIO-VILCAD score but still significantly improved risk classification (net reclassification improvement 12.5%; P < 0.001). CONCLUSION The VILCAD biomarker score based on routine parameters complemented by novel biomarkers outperforms previous risk algorithms and allows more accurate classification of patients with stable CAD, enabling physicians to choose more personalized treatment regimens for their patients.

Neural correlates of anticipation risk reflect risk preferences

Individual risk preferences have a large influence on decisions, such as financial investments, career and health choices, or gambling. Decision making under risk has been studied both behaviorally and on a neural level. It remains unclear, however, how risk attitudes are encoded and integrated with choice. Here, we investigate how risk preferences are reflected in neural regions known to process risk. We collected functional magnetic resonance images of 56 human subjects during a gambling task (Preuschoff et al., 2006). Subjects were grouped into risk averters and risk seekers according to the risk preferences they revealed in a separate lottery task. We found that during the anticipation of high-risk gambles, risk averters show stronger responses in ventral striatum and anterior insula compared to risk seekers. In addition, risk prediction error signals in anterior insula, inferior frontal gyrus, and anterior cingulate indicate that risk averters do not dissociate properly between gambles that are more or less risky than expected. We suggest this may result in a general overestimation of prospective risk and lead to risk avoidance behavior. This is the first study to show that behavioral risk preferences are reflected in the passive evaluation of risky situations. The results have implications on public policies in the financial and health domain.

Prediction of arrhythmic events by Wedensky modulation in patients with coronary artery disease.

PRINCIPLES Prediction of arrhythmic events (AEs) has gained importance with the availability of implantable cardioverter-defibrillators (ICDs), but is still imprecise. This study evaluated the innovative Wedensky modulation index (WMI) as predictor of AEs. METHODS In this prospective cohort, 179 patients with coronary artery disease (CAD) referred for AE risk assessment underwent baseline evaluation including measurement of R-/T-wave WMI (WMI(RT)) and left ventricular ejection fraction (LVEF). Two endpoints were assessed 3 years after the baseline evaluation: sudden cardiac death or appropriate ICD event (EP1) and any cardiac death or appropriate ICD event (EP2). Associations between baseline predictors (WMI(RT) and LVEF) and endpoints were evaluated in regression models. RESULTS Only three patients were lost to follow-up. EP1 and EP2 occurred in 24 and 27 patients, respectively. WMI(RT) (odds ratio [OR] per 1 point increase for EP1 20.1, 95% confidence interval [CI] 1.8-221.4, p = 0.014, and for EP2 73.3, 95% CI 6.6-817.7, p <0.001) and LVEF (OR per 1% increase for EP1 0.94, 95% CI 0.90-0.99, p = 0.013, and for EP2 0.93, 95% CI 0.89-0.97, p = 0.002) were significantly associated with both endpoints. In bivariable regression controlled for LVEF, WMI(RT) was independently associated with EP1 (p = 0.047) and EP2 (p = 0.007). The combination of WMI(RT) ≥0.60 and LVEF ≤30% resulted in a positive predictive value of 36% for EP1 and 50% for EP2. CONCLUSIONS WMI(RT) is a significant predictor of AEs independent of LVEF and has potential to improve AE risk prediction in CAD patients. However, WMI(RT) should be evaluated in larger and independent samples before recommendations for clinical routine can be made.

Normal Thyroid Function and the Risk of Atrial Fibrillation: the Rotterdam Study

CONTEXT Hyperthyroidism is an established risk factor for atrial fibrillation (AF), but information concerning the association with variations within the normal range of thyroid function and subgroups at risk is lacking. OBJECTIVE This study aimed to investigate the association between normal thyroid function and AF prospectively and explore potential differential risk patterns. DESIGN, SETTING, AND PARTICIPANTS From the Rotterdam Study we included 9166 participants ≥ 45 y with TSH and/or free T4 (FT4) measurements and AF assessment (1997-2012 median followup, 6.8 y), with 399 prevalent and 403 incident AF cases. MAIN OUTCOME MEASURES Outcome measures were 3-fold: 1) hazard ratios (HRs) for the risk of incident AF by Cox proportional-hazards models, 2) 10-year absolute risks taking competing risk of death into account, and 3) discrimination ability of adding FT4 to the CHARGE-AF simple model, an established prediction model for AF. RESULTS Higher FT4 levels were associated with higher risks of AF (HR 1.63, 95% confidence interval, 1.19-2.22), when comparing those in the highest quartile to those in lowest quartile. Absolute 10-year risks increased with higher FT4 in participants ≤ 65 y from 1-9% and from 6-12% in subjects ≥ 65 y. Discrimination of the prediction model improved when adding FT4 to the simple model (c-statistic, 0.722 vs 0.729; P = .039). TSH levels were not associated with AF. CONCLUSIONS There is an increased risk of AF with higher FT4 levels within the normal range, especially in younger subjects. Adding FT4 to the simple model slightly improved discrimination of risk prediction.

Contribution of activity and skin temperature to the risk of developing pressure ulcers in nursing facility residents

Purpose. The central concepts in pressure ulcer risk are exposure to external pressure caused by inactivity and tissue tolerance to pressure, a factor closely related to blood flow. Inactivity measures are effective in predicting pressure ulcer risk. The purpose of the study is to evaluate whether a physiological measure of skin blood flow improves pressure ulcer risk prediction. Skin temperature regularity and self-similarity, as proxy measures of blood flow, and not previously described, may be undefined pressure ulcer risk factors. The specific aims were to determine whether a sample of nursing facility residents at high risk of pressure ulcers classified using the Braden Scale for Pressure Sore Risk© differ from a sample of low risk residents according to (1) exposure to external pressure as measured by resident activity, (2) tissue tolerance to external pressure as measured by skin temperature, and (3) skin temperature fluctuations and recovery in response to a commonly occurring stressor, bathing and additionally whether (4) scores on the Braden Scale mobility subscale score are related to entropy and the spectral exponent. ^ Methods. A two group observational time series design was used to describe activity and skin temperature regularity and self-similarity, calculating entropy and the spectral exponent using detrended fluctuation analysis respectively. Twenty nursing facility residents wore activity and skin temperature monitors for one week. One bathing episode was observed as a commonly occurring stressor for skin temperature.^ Results. Skin temperature multiscale entropy (MSE), F(1, 17) = 5.55, p = .031, the skin temperature spectral exponent, F(1, 17) = 6.19, p = .023, and the activity mean MSE, F(1, 18) = 4.52, p = .048 differentiated the risk groups. The change in skin temperature entropy during bathing was significant, t(16) = 2.55, p = .021, (95% CI, .04-.40). Multiscale entropy for skin temperature was lowest in those who developed pressure ulcers, F(1, 18) = 35.14, p < .001.^ Conclusions. This study supports the tissue tolerance component of the Braden and Bergstrom conceptual framework and shows differences in skin temperature multiscale entropy between pressure ulcer risk categories, pressure ulcer outcome, and during a commonly occurring stressor. ^

Improving risk stratification in non-ST-segment elevation myocardial infarction with combined assessment of GRACE and CRUSADE risk scores

BACKGROUND: Risk assessment is fundamental in the management of acute coronary syndromes (ACS), enabling estimation of prognosis. AIMS: To evaluate whether the combined use of GRACE and CRUSADE risk stratification schemes in patients with myocardial infarction outperforms each of the scores individually in terms of mortality and haemorrhagic risk prediction. METHODS: Observational retrospective single-centre cohort study including 566 consecutive patients admitted for non-ST-segment elevation myocardial infarction. The CRUSADE model increased GRACE discriminatory performance in predicting all-cause mortality, ascertained by Cox regression, demonstrating CRUSADE independent and additive predictive value, which was sustained throughout follow-up. The cohort was divided into four different subgroups: G1 (GRACE<141; CRUSADE<41); G2 (GRACE<141; CRUSADE≥41); G3 (GRACE≥141; CRUSADE<41); G4 (GRACE≥141; CRUSADE≥41). RESULTS: Outcomes and variables estimating clinical severity, such as admission Killip-Kimbal class and left ventricular systolic dysfunction, deteriorated progressively throughout the subgroups (G1 to G4). Survival analysis differentiated three risk strata (G1, lowest risk; G2 and G3, intermediate risk; G4, highest risk). The GRACE+CRUSADE model revealed higher prognostic performance (area under the curve [AUC] 0.76) than GRACE alone (AUC 0.70) for mortality prediction, further confirmed by the integrated discrimination improvement index. Moreover, GRACE+CRUSADE combined risk assessment seemed to be valuable in delineating bleeding risk in this setting, identifying G4 as a very high-risk subgroup (hazard ratio 3.5; P<0.001). CONCLUSIONS: Combined risk stratification with GRACE and CRUSADE scores can improve the individual discriminatory power of GRACE and CRUSADE models in the prediction of all-cause mortality and bleeding. This combined assessment is a practical approach that is potentially advantageous in treatment decision-making.

Exploring the potential role of allostatic load biomarkers in risk assessment of patients presenting with depressive symptoms

Background: Depression is a major health problem worldwide and the majority of patients presenting with depressive symptoms are managed in primary care. Current approaches for assessing depressive symptoms in primary care are not accurate in predicting future clinical outcomes, which may potentially lead to over or under treatment. The Allostatic Load (AL) theory suggests that by measuring multi-system biomarker levels as a proxy of measuring multi-system physiological dysregulation, it is possible to identify individuals at risk of having adverse health outcomes at a prodromal stage. Allostatic Index (AI) score, calculated by applying statistical formulations to different multi-system biomarkers, have been associated with depressive symptoms. Aims and Objectives: To test the hypothesis, that a combination of allostatic load (AL) biomarkers will form a predictive algorithm in defining clinically meaningful outcomes in a population of patients presenting with depressive symptoms. The key objectives were: 1. To explore the relationship between various allostatic load biomarkers and prevalence of depressive symptoms in patients, especially in patients diagnosed with three common cardiometabolic diseases (Coronary Heart Disease (CHD), Diabetes and Stroke). 2 To explore whether allostatic load biomarkers predict clinical outcomes in patients with depressive symptoms, especially in patients with three common cardiometabolic diseases (CHD, Diabetes and Stroke). 3 To develop a predictive tool to identify individuals with depressive symptoms at highest risk of adverse clinical outcomes. Methods: Datasets used: ‘DepChron’ was a dataset of 35,537 patients with existing cardiometabolic disease collected as a part of routine clinical practice. ‘Psobid’ was a research data source containing health related information from 666 participants recruited from the general population. The clinical outcomes for 3 both datasets were studied using electronic data linkage to hospital and mortality health records, undertaken by Information Services Division, Scotland. Cross-sectional associations between allostatic load biomarkers calculated at baseline, with clinical severity of depression assessed by a symptom score, were assessed using logistic and linear regression models in both datasets. Cox’s proportional hazards survival analysis models were used to assess the relationship of allostatic load biomarkers at baseline and the risk of adverse physical health outcomes at follow-up, in patients with depressive symptoms. The possibility of interaction between depressive symptoms and allostatic load biomarkers in risk prediction of adverse clinical outcomes was studied using the analysis of variance (ANOVA) test. Finally, the value of constructing a risk scoring scale using patient demographics and allostatic load biomarkers for predicting adverse outcomes in depressed patients was investigated using clinical risk prediction modelling and Area Under Curve (AUC) statistics. Key Results: Literature Review Findings. The literature review showed that twelve blood based peripheral biomarkers were statistically significant in predicting six different clinical outcomes in participants with depressive symptoms. Outcomes related to both mental health (depressive symptoms) and physical health were statistically associated with pre-treatment levels of peripheral biomarkers; however only two studies investigated outcomes related to physical health. Cross-sectional Analysis Findings: In DepChron, dysregulation of individual allostatic biomarkers (mainly cardiometabolic) were found to have a non-linear association with increased probability of co-morbid depressive symptoms (as assessed by Hospital Anxiety and Depression Score HADS-D≥8). A composite AI score constructed using five biomarkers did not lead to any improvement in the observed strength of the association. In Psobid, BMI was found to have a significant cross-sectional association with the probability of depressive symptoms (assessed by General Health Questionnaire GHQ-28≥5). BMI, triglycerides, highly sensitive C - reactive 4 protein (CRP) and High Density Lipoprotein-HDL cholesterol were found to have a significant cross-sectional relationship with the continuous measure of GHQ-28. A composite AI score constructed using 12 biomarkers did not show a significant association with depressive symptoms among Psobid participants. Longitudinal Analysis Findings: In DepChron, three clinical outcomes were studied over four years: all-cause death, all-cause hospital admissions and composite major adverse cardiovascular outcome-MACE (cardiovascular death or admission due to MI/stroke/HF). Presence of depressive symptoms and composite AI score calculated using mainly peripheral cardiometabolic biomarkers was found to have a significant association with all three clinical outcomes over the following four years in DepChron patients. There was no evidence of an interaction between AI score and presence of depressive symptoms in risk prediction of any of the three clinical outcomes. There was a statistically significant interaction noted between SBP and depressive symptoms in risk prediction of major adverse cardiovascular outcome, and also between HbA1c and depressive symptoms in risk prediction of all-cause mortality for patients with diabetes. In Psobid, depressive symptoms (assessed by GHQ-28≥5) did not have a statistically significant association with any of the four outcomes under study at seven years: all cause death, all cause hospital admission, MACE and incidence of new cancer. A composite AI score at baseline had a significant association with the risk of MACE at seven years, after adjusting for confounders. A continuous measure of IL-6 observed at baseline had a significant association with the risk of three clinical outcomes- all-cause mortality, all-cause hospital admissions and major adverse cardiovascular event. Raised total cholesterol at baseline was associated with lower risk of all-cause death at seven years while raised waist hip ratio- WHR at baseline was associated with higher risk of MACE at seven years among Psobid participants. There was no significant interaction between depressive symptoms and peripheral biomarkers (individual or combined) in risk prediction of any of the four clinical outcomes under consideration. Risk Scoring System Development: In the DepChron cohort, a scoring system was constructed based on eight baseline demographic and clinical variables to predict the risk of MACE over four years. The AUC value for the risk scoring system was modest at 56.7% (95% CI 55.6 to 57.5%). In Psobid, it was not possible to perform this analysis due to the low event rate observed for the clinical outcomes. Conclusion: Individual peripheral biomarkers were found to have a cross-sectional association with depressive symptoms both in patients with cardiometabolic disease and middle-aged participants recruited from the general population. AI score calculated with different statistical formulations was of no greater benefit in predicting concurrent depressive symptoms or clinical outcomes at follow-up, over and above its individual constituent biomarkers, in either patient cohort. SBP had a significant interaction with depressive symptoms in predicting cardiovascular events in patients with cardiometabolic disease; HbA1c had a significant interaction with depressive symptoms in predicting all-cause mortality in patients with diabetes. Peripheral biomarkers may have a role in predicting clinical outcomes in patients with depressive symptoms, especially for those with existing cardiometabolic disease, and this merits further investigation.