WE-G-217A-04: Segmenting Limbic Circuit Fiber Tracts for Quantifying Changes in Diffusion Indices Due to Partial Radiation Therapy.

PURPOSE: To segment the cingulum, fornix and the corpus callosum using diffusion weighted images, in order to evaluate the radiation-induced damage on these important white matter structures Methods: We propose a seed-based tractography algorithm for segmenting the entire cingulum into the left and right anterior, superior, and posterior subdivisions. In that algorithm, multiple ROIs (regions of interest) for tractography are automatically created from seeded points by a level-set segmentation algorithm. Moreover, we propose new ROI placements for fornix tractography. Furthermore, a new subdividing method is proposed for supporting the possible rotation in the corpus callosum longitudinal axis in presence of glioma tumors. To test the reproducibility of the proposed algorithm for segmenting the cingulum, test-retest diffusion tensor data-sets of twelve patients were chosen from the National Biomedical Imaging Archive database. We calculated the Dice coefficients of the volumetric overlap between the test and retest cingulum segments. Twenty-two patients with low grade brain gliomas underwent 6-week fractionated radiation therapy (RT). The patients prospectively enrolled MRI with diffusion tensor imaging up to 18 months after starting RT. The fornix and subdivisions of cingulum and corpus callosum were segmented. RESULTS: The Dice coefficients between test and retest data were 0.878, 0.899, 0.951, 0.954, 0.926, and 0.929 for the anterior right and left, the superior right and left, and the posterior right and left cingulum segments respectively, suggesting our method's high reproducibility. Amongst the fornix and subdivisions of cingulum and corpus callosum, the most significant changes in fractional anisotropy, mean, axial, and radial diffusivity were observed in fornix throughout all dose volumes, indicating that both axonal degradation and dysmyelination are prominent radiation effects in the fornix. CONCLUSIONS: The proposed seed-based tractography method segments the cingulum accurately and completely with dramatically reducing the operator's interaction time and effort in manual depicting large number of ROIs. Supported in part by National Institutes of Health grants RO1 NS064973.

Metabolic alterations: a biomarker for radiation-induced normal brain injury-an MR spectroscopy study.

PURPOSE: To assess if interval changes in metabolic status in normal cerebral tissue after radiation therapy (RT) can be detected by 2D CSI (chemical shift imaging) proton spectroscopy. MATERIALS AND METHODS: Eleven patients with primary brain tumors undergoing cranial radiation therapy (RT) were included. 2D-CSI MRS was performed before, during, and after the course of RT with the following parameters: TE/TR 144/1500 ms, field of view (FOV) 24, thickness 10 mm, matrix 16 x 16. The metabolic ratios choline/creatine (Cho/Cr), N-acetylaspartate (NAA)/Cr, and NAA/Cho in normal brain tissue were calculated. RESULTS: NAA/Cr and Cho/Cr were significantly decreased at week 3 during RT and at 1 month and 6 months after RT compared to values prior to RT (P < 0.01). The NAA/Cr ratio decreased by -0.19 +/- 0.05 (mean +/- standard error [SE]) at week 3 of RT, -0.14 +/- 0.06 at the last week of RT, -0.14 +/- 0.05 at 1 month after RT, and -0.30 +/- 0.08 at 6 months after RT compared to the pre-RT value of 1.43 +/- 0.04. The Cho/Cr ratio decreased by -0.27 +/- 0.05 at week 3 of RT, -0.11 +/- 0.05 at the last week of RT, -0.26 +/- 0.05 at 1 month after RT and -0.25 +/- 0.07 at 6 months after RT from the pre-RT value of 1.29 +/- 0.03. Changes in Cho/Cr were correlated with the interaction of the radiation dose and dose-volume at week 3 of RT, during the last week of RT (P < 0.005), and at 1 month after RT (P = 0.017). CONCLUSION: The results of this study suggest that MRS can detect early metabolic changes in normal irradiated brain tissue.

CHIP TUNER: a web tool for evidence-based noise reduction in gene discovery.

The potential for gene discovery, fueled by DNA microchip technology and the sequencing of hundreds of genomes, is unprecedented. In this context, trying to discover genes that are actually of significance rather than merely appearing so due to noise is of utmost importance. We present a web application, CHIP TUNER, which assists in this gene discovery process. Our system uses evidence-based noise reduction to help delineate candidate target genes of biological importance. Specifically, CHIP TUNER learns from redundant experiments an "identity mask" that defines a region of noise inherent to biological sampling and DNA microarray processing; it then takes this into account during actual sample comparisons. The goal of CHIP TUNER is to improve the chances that newly discovered "important" genes are actually of importance before large amounts of time and resources are invested.

Stereotactic radiosurgery in the management of angiographically occult vascular malformations.

PURPOSE: To evaluate the role of stereotactic radiosurgery in the treatment of angiographically occult vascular malformations (AOVMs). METHODS AND MATERIALS: From 1987 to 1996, 21 patients, 10 males and 11 females, median age of 41 years (range: 7-75 years), with an intracerebral AOVM underwent stereotactic radiosurgery at our institution. All were considered at high risk for surgical intervention. The vascular lesions were located in the brainstem (17 patients), basal ganglia (2), occipital lobe (1), and cerebellum (1). Diagnosis was based on high-resolution magnetic resonance imaging (MRI). Clinical presentation at onset included previous intracerebral hemorrhage (20 patients) and epilepsy (1). All patients were treated with a linac-based radiosurgical technique. The median dose delivered was 25 Gy (range 13-50 Gy), typically prescribed to the 80-90% isodose surface (range 50-90%), which corresponded to the periphery of the vascular malformation. Patients were followed by clinical neurologic assessment and by MRI on a regular interval basis. RESULTS: Follow-up was obtained in 20 patients; clinical or MRI information was not available for 1 patient, and this patient was excluded from our analysis. At a median follow-up of 77 months (range: 4-141 months), follow-up MRIs postradiosurgery do not demonstrate any changes in the appearance of the AOVM. Four patients developed an intracranial bleed at 4, 8, 35, and 57 months postradiosurgery. Annual hemorrhage rates were considerably higher in the observation period preradiosurgery than postradiosurgery (30% vs. 3.2%, p < 0.001). Complications postradiosurgery were observed in 4 patients. Three patients developed mild to moderate edema surrounding the radiosurgical target, expressed at 5, 8, and 24 months, respectively. In all cases, the edema was transient and resolved completely on subsequent MRIs. One of the 4 patients developed radiation necrosis 8 months after radiosurgery. CONCLUSION: The use of stereotactic radiosurgery in the treatment of AOVM continues to be controversial. Our results appear to show a reduction in the risk of symptomatic hemorrhage post treatment. Patients with previous history of hemorrhage or progressive neurologic deficit and small, well circumscribed lesions may benefit from a trial of stereotactic radiosurgery.

On reporting fold differences.

As we enter an age in which genomics and bioinformatics make possible the discovery of new knowledge about the biological characteristics of an organism, it is critical that we attempt to report newly discovered "significant" phenotypes only when they are actually of significance. With the relative youth of genome-scale gene expression technologies, how to make such distinctions has yet to be better defined. We present a "mask technology" by which to filter out those levels of gene expression that fall within the noise of the experimental techniques being employed. Conversely, our technique can lend validation to significant fold differences in expression level even when the fold value may appear quite small (e.g. 1.3). Given array-organized expression level results from a pair of identical experiments, our ID Mask Tool enables the automated creation of a two-dimensional "region of insignificance" that can then be used with subsequent data analyses. Fundamentally, this should enable researchers to report on findings that are more likely to be in nature truly meaningful. Moreover, this can prevent major investments of time, energy, and biological resources into the pursuit of candidate genes that represent false positives.

Building ICU artifact detection models with more data in less time.

As many as 86% of intensive care unit (ICU) alarms are false. Multiple signal integration of temporal monitor data by decision tree induction may improve artifact detection. We explore the effect of data granularity on model-building by comparing models made from 1-second versus 1-minute data. Models developed from 1-minute data remained effective when tested on 1-second data. Model development using 1-minute data means that more hours of ICU monitoring (including more artifacts) can be processed in less time. Compression of temporal data by arithmetic mean, therefore, can be an effective method for decreasing knowledge discovery processing time without compromising learning.

Event discovery in medical time-series data.

Vast amounts of clinical information are generated daily on patients in the health care setting. Increasingly, this information is collected and stored for its potential utility in advancing health care. Knowledge-based systems, for example, might be able to apply rules to the collected data to determine whether a patient has a certain condition. Often, however, the underlying knowledge needed to write such rules is not well understood. How could these clinical data be useful then? Use of machine learning is one answer. We present a pipeline for discovering the knowledge needed for event detection in medical time-series data. We demonstrate how this process can be applied in the development of intelligent patient monitoring for the intensive care unit (ICU). Specifically, we develop a system for detecting Otrue alarmO situations in the ICU, where currently as many as 86% of bedside monitor alarms are false.

Multiple signal integration by decision tree induction to detect artifacts in the neonatal intensive care unit.

The high incidence of false alarms in the intensive care unit (ICU) necessitates the development of improved alarming techniques. This study aimed to detect artifact patterns across multiple physiologic data signals from a neonatal ICU using decision tree induction. Approximately 200 h of bedside data were analyzed. Artifacts in the data streams were visually located and annotated retrospectively by an experienced clinician. Derived values were calculated for successively overlapping time intervals of raw values, and then used as feature attributes for the induction of models trying to classify 'artifact' versus 'not artifact' cases. The results are very promising, indicating that integration of multiple signals by applying a classification system to sets of values derived from physiologic data streams may be a viable approach to detecting artifacts in neonatal ICU data.

Using classification tree and logistic regression methods to diagnose myocardial infarction.

Early and accurate diagnosis of myocardial infarction (MI) in patients who present to the Emergency Room (ER) complaining of chest pain is an important problem in emergency medicine. A number of decision aids have been developed to assist with this problem but have not achieved general use. Machine learning techniques, including classification tree and logistic regression (LR) methods, have the potential to create simple but accurate decision aids. Both a classification tree (FT Tree) and an LR model (FT LR) have been developed to predict the probability that a patient with chest pain is having an MI based solely upon data available at time of presentation to the ER. Training data came from a data set collected in Edinburgh, Scotland. Each model was then tested on a separate Edinburgh data set, as well as on a data set from a different hospital in Sheffield, England. Previously published models, the Goldman classification tree[1] and Kennedy LR equation[2], were evaluated on the same test data sets. On the Edinburgh test set, results showed that the FT Tree, FT LR, and Kennedy LR performed equally well, with ROC curve areas of 94.04%, 94.28%, and 94.30%, respectively, while the Goldman Tree's performance was significantly poorer, with an area of 84.03%. The difference in ROC areas between the first three models and the Goldman model is significant beyond the 0.0001 level. On the Sheffield test set, results showed that the FT Tree, FT LR, and Kennedy LR ROC areas were not significantly different (p > = 0.17), while the FT Tree again outperformed the Goldman Tree (p = 0.006). Unlike previous work[3], this study indicates that classification trees, which have certain advantages over LR models, may perform as well as LR models in the diagnosis of patients with MI.

Poor prognosis for existing monitors in the intensive care unit.

OBJECTIVE: To identify areas requiring the most urgent improvement in the intensive care unit (ICU); and to accurately determine the positive predictive value of routine critical care patient monitoring alarms, as well as the common causes for false-positive alarms. DESIGN: Prospective, observational study. SETTING: A multidisciplinary ICU in a university-affiliated children's hospital (excluding children with primary heart disease). INTERVENTIONS: The occurrence rate, cause, and appropriateness of all alarms from tracked monitors were recorded by a trained observer and validated by the bedside nurse over a 10-wk period for a single bedspace at a time. MEASUREMENTS AND MAIN RESULTS: After 298 monitored hrs, 86% of a total 2,942 alarms were found to be false-positive alarms, while an additional 6% were classified as clinically irrelevant true alarms. Only 8% of all alarms tracked during the study period were determined to be true alarms with clinical significance. Alarms were also classified according to whether they were clearly associated with a "patient intervention" (18%), were clearly not associated with a patient intervention (74%), or had unclear association to interventions (8%). While 11% of "nonpatient intervention" alarms were clinically significant true alarms, only 2% of "patient intervention" alarms were so. Positive predictive values for the various devices ranged from < 1% for the pulse oximeter's heart rate signal to 74% for the arterial catheter's mean systemic blood pressure signal during periods free from patient interventions. The pulse oximeter caused false-positive alarms most frequently, with common reasons being bad data format/bad connection and poor contact. CONCLUSION: Efforts to develop intelligent monitoring systems have more potential to deliver significantly improved patient care by initially targeting especially weak areas in ICU monitoring, such as pulse oximetry reliability.