Probability machines: consistent probability estimation using nonparametric learning machines.

BACKGROUND: Most machine learning approaches only provide a classification for binary responses. However, probabilities are required for risk estimation using individual patient characteristics. It has been shown recently that every statistical learning machine known to be consistent for a nonparametric regression problem is a probability machine that is provably consistent for this estimation problem. OBJECTIVES: The aim of this paper is to show how random forests and nearest neighbors can be used for consistent estimation of individual probabilities. METHODS: Two random forest algorithms and two nearest neighbor algorithms are described in detail for estimation of individual probabilities. We discuss the consistency of random forests, nearest neighbors and other learning machines in detail. We conduct a simulation study to illustrate the validity of the methods. We exemplify the algorithms by analyzing two well-known data sets on the diagnosis of appendicitis and the diagnosis of diabetes in Pima Indians. RESULTS: Simulations demonstrate the validity of the method. With the real data application, we show the accuracy and practicality of this approach. We provide sample code from R packages in which the probability estimation is already available. This means that all calculations can be performed using existing software. CONCLUSIONS: Random forest algorithms as well as nearest neighbor approaches are valid machine learning methods for estimating individual probabilities for binary responses. Freely available implementations are available in R and may be used for applications.

Patient-centered yes/no prognosis using learning machines.

In the last 15 years several machine learning approaches have been developed for classification and regression. In an intuitive manner we introduce the main ideas of classification and regression trees, support vector machines, bagging, boosting and random forests. We discuss differences in the use of machine learning in the biomedical community and the computer sciences. We propose methods for comparing machines on a sound statistical basis. Data from the German Stroke Study Collaboration is used for illustration. We compare the results from learning machines to those obtained by a published logistic regression and discuss similarities and differences.

Practical experiences on the necessity of external validation.

The validity of prognostic models is an important prerequisite for their applicability in practical clinical settings. Here, we report on a specific prognostic study on stroke patients and describe how we explored the prediction performance of our model. We considered two practically highly relevant generalization aspects, namely, the model's performance in patients recruited at a later time point (temporal transportability) and in medical centers different from those used for model building (geographic transportability). To estimate the accuracy of the model, we investigated classical internal validation techniques and leave-one-center-out cross validation (CV). Prognostic models predicting functional independence of stroke patients were developed in a training set using logistic regression, support vector machines, and random forests (RFs). Tenfold CV and leave-one-center-out CV were employed to estimate temporal and geographic transportability of the models. For temporal and external validation, the resulting models were used to classify patients from a later time point and from different clinics. When applying the regression model or the RFs, accuracy in the temporal validation data was well predicted from classical internal validation. However, when predicting geographic transportability all approaches had difficulties. We observed that the leave-one-center-out CV yielded better estimates than classical CV. On the basis of our results, we conclude that external validation in patients from different clinics is required before a prognostic model can be applied in practice. Even validating the model in patients recruited merely at a later time point does not suffice to predict how it may fare with regard to another clinic.

A kinetic analysis of calcium-triggered exocytosis.

Although the relationship between exocytosis and calcium is fundamental both to synaptic and nonneuronal secretory function, analysis is problematic because of the temporal and spatial properties of calcium, and the fact that vesicle transport, priming, retrieval, and recycling are coupled. By analyzing the kinetics of sea urchin egg secretory vesicle exocytosis in vitro, the final steps of exocytosis are resolved. These steps are modeled as a three-state system: activated, committed, and fused, where interstate transitions are given by the probabilities that an active fusion complex commits (alpha) and that a committed fusion complex results in fusion, p. The number of committed complexes per vesicle docking site is Poisson distributed with mean n. Experimentally, p and n increase with increasing calcium, whereas alpha and the pn ratio remain constant, reducing the kinetic description to only one calcium-dependent, controlling variable, n. On average, the calcium dependence of the maximum rate (R(max)) and the time to reach R(max) (T(peak)) are described by the calcium dependence of n. Thus, the nonlinear relationship between the free calcium concentration and the rate of exocytosis can be explained solely by the calcium dependence of the distribution of fusion complexes at vesicle docking sites.

Studies of murine schistosomiasis reveal interleukin-13 blockade as a treatment for established and progressive liver fibrosis.

In several allergic, autoimmune, and infectious diseases, fibrosis is a major cause of morbidity and mortality. Here, using a model of infection-induced liver fibrosis, we show that interleukin (IL)-13 is required at all stages of Schistosomiasis mansoni infection to induce fibrosis. IL-4 production was preserved in IL-13-deficient mice, yet failed to significantly contribute to the fibrotic response in either acute or chronic infection. Significant fibrosis develops in all infected mice, although the magnitude of the response varies widely in inbred mice. C3H/HeN, BALB/c, and C57BL/6 mice develop high, intermediate, and low levels of fibrosis, respectively. Despite these differences, IL-13 antagonism resulted in a marked amelioration of fibrosis in all strains. The fibrotic mechanism in the high- and low-responder strains was unrelated to their tissue eosinophil or mast cell responses, but did correlate with their patterns of IL-13, IL-10, and interferon gamma (IFN-gamma) mRNA expression. Indeed, severe fibrosis correlated with a high IL-13 and low IFN-gamma/IL-10 mRNA response. Because fibrotic diseases are typically progressive disorders, an important issue was to determine whether IL-13 inactivation might be used to treat an established and ongoing fibrotic disease. Here, IL-13 antagonism was highly efficacious, even after fibrosis and the Th2 cytokine response were firmly established. These studies demonstrate the central role played by IL-13 in fibrogenesis and suggest that therapeutic approaches aimed at disrupting the IL-13 pathway will be highly effective at preventing fibrotic disease caused by chronic Th2-mediated inflammatory reactions.

Automated polyp detection at CT colonography: feasibility assessment in a human population.

PURPOSE: To test the feasibility of and improve a computer algorithm to automatically detect colonic polyps in real human computed tomographic (CT) colonographic data sets. MATERIALS AND METHODS: Twenty patients with known polyps underwent CT colonography in the supine position. CT colonographic data were processed by using a shape-based algorithm that depicts masses that protrude into the lumen. We studied nine shape criteria and three isosurface threshold settings. Results were compared with those of conventional colonoscopy performed the same day. RESULTS: There were 50 polyps (28 were > or =10 mm in size; 12, 5-9 mm; 10, <5 mm). The sensitivity with optimal settings for detecting polyps 10 mm or greater was 64% (18 of 28). Sensitivity improved to 71% (10 of 14) for polyps 10 mm or greater in well-distended colonic segments. Performance decreased for polyps less than 10 mm, poorly distended colonic segments, and other shape algorithms. There was a mean of six false-positive lesion sites per colon. These sites were reduced 39% to 3.5 per colon by sampling CT attenuation at the lesion site and discarding sites having attenuation less than a threshold. CONCLUSION: Automated detection of colonic polyps, especially clinically important large polyps, is feasible. Colonic distention is an important determinant of sensitivity. Further increases in sensitivity may be achieved by adding prone CT colonography.

Comparison of novel and existing methods for detection of linkage disequilibrium using parent-child trios in the GAW12 genetic isolate simulated data.

A novel method for joint detection of association caused by linkage disequilibrium (LD) and estimation of both recombination fraction and linkage disequilibrium parameters was compared to several existing implementations of the transmission/disequilibrium test (TDT) and modifications of the TDT in the simulated genetic isolate data from Genetic Analysis Workshop 12. The first completely genotyped trio of affected child and parents was selected from each family in each replicate so that the TDT tests are valid tests of linkage and association, rather than being only valid as tests for linkage. In general, power to detect LD using the genome-wide scan markers was inadequate in the individual replicate samples, but the power was better when analyzing several SNP markers in candidate gene 1.

Evaluating UMLS strings for natural language processing.

The National Library of Medicine's Unified Medical Language System (UMLS) is a rich source of knowledge in the biomedical domain. The UMLS is used for research and development in a range of different applications, including natural language processing (NLP). In this paper we investigate the nature of the strings found in the UMLS Metathesaurus and evaluate them for their usefulness in NLP. We begin by identifying a number of properties that might allow us to predict the likelihood of a given string being found or not found in a corpus. We use a statistical model to test these predictors against our corpus, which is derived from the MEDLINE database. For one set of properties the model correctly predicted 77% of the strings that do not belong to the corpus, and 85% of the strings that do belong to the corpus. For another set of properties the model correctly predicted 96% of the strings that do not belong to the corpus and 29% of the strings that do belong to the corpus.

In vivo microdialysis assessment of extracellular serotonin and dopamine levels in awake monkeys during sustained fluoxetine administration.

Fluoxetine (FLU) rapidly enhances extracellular (EC) serotonin (5-HT) in rodent brain, whereas the antidepressant effects of this drug in humans are typically not observed for 2-3 weeks. Thus, the effects of chronic oral FLU administration on neocortical and hippocampal EC 5-HT, and on caudate EC 5-HT and dopamine (DA), were examined in awake monkeys (Macaca fascicularis) using in vivo microdialysis (10.0 mg/kg; 3, 7, 14, and 21 days). On day 3, 5-HT was significantly increased above baseline levels in hippocampus (HC) and caudate. There was a trend for an increase in neocortex EC 5-HT levels. However, by day 7 5-HT remained significantly elevated only in HC, although 5-HT levels elsewhere had not completely returned to baseline. In contrast, levels of the 5-HT metabolite, 5-HIAA, were significantly reduced in all brain regions at all time points. Caudate DA levels tended to be decreased throughout FLU treatment. Local FLU and K(+) infusion were also used at various times during chronic systemic FLU administration to evaluate changes in functional synaptic regulation. In general, these results, along with the significant decrease in 5-HIAA levels and the tendency for basal EC 5-HT levels to remain modestly elevated only in HC during sustained FLU administration, suggest a reduction in releasable pools of 5-HT. Taken together with the trend for a decrease in caudate EC DA levels, these results do not appear to support the current hypothesis regarding the mechanism of action of SSRI antidepressants-that monoaminergic neurotransmission is progressively augmented during chronic treatment.

Automated polyp detector for CT colonography: feasibility study.

An abdominal computed tomographic scan was modified by inserting 10 simulated colonic polyps with use of methods that closely mimic the attenuation, noise, and polyp-colon wall interface of naturally occurring polyps. A shape-based polyp detector successfully located six of the 10 polyps. When settings that enhanced the edge profile of polyps were chosen, eight of 10 polyps were detected. There were no false-positive detections. Shape analysis is technically feasible and is a promising approach to automated polyp detection.

Many multipotential gene-marked progenitor or stem cell clones contribute to hematopoiesis in nonhuman primates.

Retroviral insertion site analysis was used to track the contribution of retrovirally transduced primitive progenitors to hematopoiesis after autologous transplantation in the rhesus macaque model. CD34-enriched mobilized peripheral blood cells were transduced with retroviral marking vectors containing the neo gene and were reinfused after total body irradiation. High-level gene transfer efficiency allowed insertion site analysis of individual myeloid and erythroid colony-forming units (CFU) and of highly purified B- and T-lymphoid populations in 2 animals. At multiple time points up to 1 year after transplantation, retroviral insertion sites were identified by performing inverse polymerase chain reaction and sequencing vector-containing CFU or more than 99% pure T- and B-cell populations. Forty-eight unique insertion sequences were detected in the first animal and also in the second animal, and multiple clones contributed to hematopoiesis at 2 or more time points. Multipotential clones contributing to myeloid and lymphoid lineages were identified. These results support the concept that hematopoiesis in large animals is polyclonal and that individual multipotential stem or progenitor cells can contribute to hematopoiesis for prolonged periods. Gene transfer to long-lived, multipotent clones is shown and is encouraging for human gene therapy applications.

Disease progression in sporadic inclusion body myositis: observations in 78 patients.

Functional decline for each decade at symptom onset and need for cane, walker, or wheelchair were assessed in 78 biopsy-proved patients with sporadic inclusion body myositis. Patients with disease onset between 40 and 59 years used a walker after 10.2 +/- 5.8 years, whereas those with disease onset between 60 and 79 years used a walker after 5.7 +/- 5.0 years (p = 0.05). Because patients progress faster to disability when symptoms begin after the age of 60, age at disease onset may define patient subsets for stratification in clinical trials.

Quantitative analysis of globin gene induction in single human erythroleukemic cells.

The mechanisms involved in the normal developmental regulation of globin gene expression, and the response to pharmacological agents that elevate fetal hemoglobin, may be expected to involve either changes in each cell or a selection process affecting subsets of differentiating erythroid cells. To study these mechanisms we have developed assays to measure mRNA levels in single erythroid cells. The assay involved the use of globin-specific probes, with no detectable cross-reactivity, in real-time, fluorescence-based quantitative PCR (Q-PCR). We had previously used this Q-PCR method to measure globin mRNA levels in cultures of primary erythroid cells demonstrating that drugs like hydroxyurea, 5-azacytidine and butyric acid each yielded increases in gamma/( gamma + ss) mRNA ratios, with differential effects on ss-globin levels. We have now extended this approach to measure globin mRNA levels in single K562 cells, a human erythroleukemic cell line, with and without 30 microM hemin treatment. Hemin exposure increases total hemoglobin levels by approximately 9-fold and total alpha-, epsilon- and gamma-globin mRNA levels by 1.5-2.3-fold. Single cell analyses showed initial wide distributions of each of the three individual globin mRNA levels with most cells having detectable but very low levels of each globin transcript. Hemin induction shifted the distributions to higher levels, with a tendency to residual left skewing as some cells remained with very low expression levels despite the effect of hemin in increasing expression in most of these low expressing cells. Thus transcriptional heterogeneity remains a crucial variable, even in this extensively used model of human erythroid biology, and clearly influences strongly the response to inducing agents. These methods may enable us to define better possible molecular and/or cellular models of globin gene modulation.

Juvenile idiopathic inflammatory myopathy: exercise-induced changes in muscle at short inversion time inversion-recovery MR imaging.

PURPOSE: To study the effect of exercise on short inversion time inversion-recovery (STIR) magnetic resonance (MR) images of thigh muscles in children with juvenile idiopathic inflammatory myopathy. MATERIALS AND METHODS: Thirty-two MR studies were performed in 19 patients with juvenile idiopathic inflammatory myopathy who performed stair-stepping exercise for up to 10 minutes (mean, 5.7 minutes). Baseline T1-weighted (n = 32) and STIR (n = 32) images and STIR images immediately (n = 32) and at 30 (n = 24) and 60 (n = 29) minutes after exercise were obtained at 0.5 T. Four radiologists graded STIR signal intensity changes, in observer performance experiments in which they were blinded to the order of image acquisition in relation to exercise. RESULTS: Changes in muscle signal intensity were observed on STIR images obtained immediately after exercise in 20 of 32 (63%) studies. The mean signal intensity score immediately after exercise (1.7 +/- 1.0 [SD]) increased compared with the mean baseline score (1.4 +/- 1.1) (P = .0005) and resolved by 30 minutes after exercise. The magnitude of exercise-induced changes correlated with the amount of work performed (r = 0.51, P = .003) but not with disease activity or baseline signal intensity when the changes were corrected for work (r < 0.17, P > .35). Radiologists demonstrated moderate to substantial agreement in the grading of signal intensity changes after exercise (kappa = 0.60-0.84). CONCLUSION: In patients with juvenile idiopathic inflammatory myopathy, stair-stepping exercise induces signal intensity changes on STIR MR studies of muscle for approximately 30 minutes after exercise, in a distribution that may mimic active muscle inflammation.

Polypoid lesions of airways: early experience with computer-assisted detection by using virtual bronchoscopy and surface curvature.

PURPOSE: To test the application of a technique developed by the authors for the computer-assisted diagnosis of polypoid airway lesions from surface rendered virtual bronchoscopic reconstructions. MATERIALS AND METHODS: A computer algorithm was developed to detect polypoid airway lesions by means of segmentation of the bronchial surface with curvature classification. This method was tested with a bronchial phantom, five cadaveric lung specimens, and virtual bronchoscopic studies in 16 patients. RESULTS: For the patient studies, the sensitivity and specificity of the method were 47%-88% and 58%-89%, respectively, depending on the value of an adjustable parameter (the mean curvature threshold). The sensitivity increased (by 20% to 34%) when only lesions larger than 5 mm in diameter were considered. CONCLUSION: With this method, polypoid airway lesions can be detected automatically, although false-positive diagnoses present an important limitation.

Recovery of reproductive function in rats treated with the aromatase inhibitor fadrozole.

The aromatase inhibitor, fadrozole hydrochloride (CGS 16949A), was developed for the treatment of breast cancer, and has not been available for pediatric use because of the lack of information about potential reproductive toxicology. To determine the effect of fadrozole on subsequent fertility and reproductive performance in rats, peripubertal male and female Sprague-Dawley rats (10/group) were given fadrozole by oral gavage once a day for 60 consecutive days (age 21 through 80 d) at a dose of 0, 1200, or 6000 micrograms/kg/d (dose range in women with breast cancer: 60 to 240 micrograms/kg/d). Following a 30-d recovery period (days 81 through 111 of age), cohabitation with untreated rats of the opposite sex was accomplished for 30 d or until positive evidence of mating was obtained (daily vaginal smears). The nonfadrozole-treated males used for cohabitation were proven fertile breeders; the females were virgin with proven 4-d estrous cycles. The duration of pregnancy, number, sex, condition, and body weight of pups were determined. Pregnant rats were weighed on gestational days 7, 14, and 20. There was a profound decrease in the number of estrous cycles at both dose levels of fadrozole compared to the control (P < 0.001). During the 30-d recovery period, estrous cycles were reestablished within a few days in the treated rats and the number and length of estrous cycles were not statistically different between fadrozole-treated and control rats. The gestational body weights of fadrozole-treated and untreated females did not differ significantly. There were no statistically significant differences in the number of matings/number of pairings, gestational length, mean live pups/litter, % pups born alive/litter, and % male pups/litter in the three groups (vehicle-, low-, and high-dose fadrozole-treated females, cohabited with untreated males and fadrozole-treated males, cohabited with untreated females). Thus, young male and female rats treated for 60 d with large doses of fadrozole had no detectable adverse effect on subsequent reproductive function.

Effects of light on T-cells in HIV-infected subjects are not dependent on history of seasonal affective disorder.

In order to evaluate the effect of light on helper- and suppressor-T-cell counts in HIV-infected individuals, with and without a history of seasonal affective disorder (SAD), we treated 35 subjects with 45 min of light therapy in the morning, in a crossover design involving two 2 week treatment conditions: visible white light (half-peak band width, 530-620 nm; 10,000 lux) and visible red light (half-peak band width, 615-685 nm; 175 lux). We found small but significant differences between the two treatment conditions, with higher CD4 and CD8 levels during the white, as compared with the red, condition. There were no differences between baseline and treatment conditions. Both light treatment conditions were associated with significant mood improvements in the SAD, but not the non-SAD, subjects. There was no evidence that the higher cell counts seen under the bright light conditions were mediated by the effects of light on mood or on plasma cortisol levels. While the size of the light effect on T-cells renders questionable the potential therapeutic value of this treatment modality for HIV, the finding is of theoretical interest and is reassuring for those HIV-infected individuals who require light treatment for other reasons.

A comparison of the standard high dose dexamethasone suppression test and the overnight 8-mg dexamethasone suppression test for the differential diagnosis of adrenocorticotropin-dependent Cushing's syndrome.

To improve the overnight 8-mg dexamethasone (DEX) suppression test (DST) for differential diagnosis of Cushing's syndrome and to develop optimal criteria for its interpretation, we increased the number of blood samples and measured the suppression of both plasma ACTH and cortisol. Forty-one patients who were subsequently proven at surgery to have Cushing's syndrome were studied (34 Cushing's disease and 7 ectopic ACTH secretion). DEX (8 mg, orally) was administered at 2300 h. Blood samples for ACTH and cortisol measurements were obtained at 0800, 0830, and 0900 h on the day before and at 0700, 0800, 0900, and 1000 h on the morning after DEX treatment. The conventional 6-day DST was also performed, with measurement of both urinary free cortisol and urinary 17-hydroxysteroids as indices of suppression. Optimal criteria for the diagnosis of Cushing's disease were developed for both the overnight 8-mg and the 6-day tests using receiver operating characteristic curves. The results were compared with those using the previously published criteria for diagnosis of Cushing's disease by the overnight 8-mg test (> 50% suppression of plasma cortisol at 0700-0800 h). In our patients, the previously published criterion for the overnight 8-mg test yielded high sensitivity (88%), but low specificity (57%), in making the diagnosis of Cushing's disease. When the time of cortisol measurement and the diagnostic criteria for Cushing's disease were revised to achieve 100% specificity, the sensitivity of the overnight 8-mg test was 71%, which was not significantly different from that of the 6-day test (79%; P = NS). Addition of plasma ACTH levels to the test did not improve diagnostic accuracy compared to that with measurement of plasma cortisol levels alone. When the revised 8-mg overnight dexamethasone suppression test was combined with the 6-day dexamethasone suppression test, sensitivity increased to 91%, with a specificity of 100%, which was significantly better than that of the overnight 8-mg test alone (P < 0.002). We conclude that the overnight 8-mg DST has low specificity for the diagnosis of Cushing's disease when performed as originally described. However, with revised sampling times and diagnostic criteria, the overnight test has sensitivity and specificity similar to those of the conventional 6-day DST. The diagnostic performance of a criterion that combines the results of both tests is better than the diagnostic performance of either test alone.

Comparison of breathing comfort during weaning with two ventilatory modes.

In twenty-one patients ventilated for > or = 3 days, we compared similar levels of partial support provided by synchronized intermittent mandatory ventilation (SIMV) and pressure support ventilation (PSV) in terms of breathing comfort. On a single day, eligible subjects experienced, in random order, both SIMV and PSV weaning protocols (sequential 20% reductions in support at timed intervals) separated by a 1 to 3 h rest. Breathing comfort was defined by subjective ratings of dyspnea and anxiety. Subjects reported significant levels of preweaning dyspnea and anxiety despite resting for at least 6 h. Dyspnea and anxiety were not significantly different between the two methods at any level of support. Our findings suggest that dyspnea and anxiety are higher than expected on "full" ventilator support, and that comfort may not differ between PSV and SIMV during active withdrawal of machine support.

Combining insulin-like growth factor-I and mean spontaneous nighttime growth hormone levels for the diagnosis of growth hormone deficiency.

There is no gold standard for the diagnosis of GH deficiency. Recent data show that spontaneous GH levels may lack sensitivity, and that GH stimulation tests lack specificity as currently performed. Serum insulin-like growth factor-I (IGF-I) measurements lack both sensitivity and specificity. Some of these problems may be explained by nutritional effects. In children, overnutrition decreases GH and increases IGF-I, while undernutrition decreases IGF-I and increases GH. To overcome these difficulties and improve diagnostic accuracy, we combined mean spontaneous nighttime GH levels with IGF-I levels in a statistically based bivariate model. On a two-dimensional plot of mean spontaneous nighttime GH level (in SD units) vs. IGF-I level (in SD units), we defined a new variable, S (sum) score, where S = (1/square root of 2) x (nighttime mean GH SD+IGF-I SD). While IGF-I (SD) and the mean spontaneous nighttime GH (SD) showed a significant correlation with body mass index, the S score was independent of body mass. We, therefore, used the S score to define a new test for GH deficiency. A child failed this bivariate test if his S score was less than -2 SD. We applied this model to 47 normal children and 48 short or slowly growing children (all prepubertal). We measured spontaneous nighttime GH levels and IGF-I levels in all children. In addition, the short children underwent 3 GH stimulation tests. Forty-six of the 47 normal children passed the bivariate test for GH sufficiency. Twenty-three of the 48 short or slowly growing children failed the bivariate test, whereas only 11 children had an abnormally low mean spontaneous nighttime GH measurement alone. Sixteen of 23 children who were GH deficient by the bivariate test were also GH deficient by the stimulation tests. In summary, the bivariate test for GH deficiency appears 1) to be independent of body mass, unlike either IGF-I or GH individually; 2) to identify more children than the mean spontaneous nighttime GH level alone; and 3) to be highly specific in the normal population, unlike stimulation tests.