Automated Machine Learning Algorithm Research Articles

P4347Automated echocardiographic quantification of left ventricular ejection fraction without volume measurements using a machine learning algorithm mimicking a human expert

Abstract Background Echocardiographic quantification of left ventricular (LV) ejection fraction (EF) relies on either manual or automated identification of endocardial boundaries followed by standard calculation of model-based end-systolic and end-diastolic LV volumes. Recent developments in artificial intelligence resulted in computer algorithms that allow near automated detection of endocardial boundaries and measurement of LV volumes and function. However, boundary identification is still prone to errors limiting accuracy in certain patients. We hypothesized that a fully automated machine learning algorithm could be developed, which circumvents border detection and instead estimates the degree of ventricular contraction, similar to a human expert trained on tens of thousands of images. Purpose This study was designed to test the feasibility and accuracy of this approach. Methods Machine learning algorithm was developed and trained on a database of >50,000 echocardiographic studies, including multiple apical 2- and 4-chamber views, to automatically estimate LVEF (AutoEF, BayLabs). Testing was performed on an independent group of 99 unselected patients, whose automated EF values were compared to reference values obtained by averaging measurements by 3 experts using conventional volume-based technique. Inter-technique agreement was assessed using linear regression and Bland-Altman analysis of bias and limits of agreement (LOA). Consistency was assessed by mean absolute deviation (MAD) among automated estimates based on different combinations of apical views. Finally, sensitivity and specificity of detecting of EF≤35% was calculated. These metrics were compared side-by-side against the same reference standard to those obtained from conventional EF measurements by clinical readers. Results Automated estimation of LVEF was feasible in all 99 patients. AutoEF values showed high consistency (MAD=2.9%) and excellent agreement with the reference values: r=0.95, bias=1.0%, LOA=±11.8%, with sensitivity 0.90 and specificity 0.92 for detection of EF≤35%. This was similar to clinicians' measurements: r=0.94, bias=1.4%, LOA=±13.4%,sensitivity 0.93, specificity 0.87. Conclusions Machine learning algorithm for volume-independent LVEF estimation is highly feasible and similar in accuracy to conventional volume-based measurements, when compared to reference values provided by an expert panel. Acknowledgement/Funding Bay Labs, Inc.

A machine learning model to predict hepatocellular carcinoma response to transcatheter arterial chemoembolization.

Some patients with hepatocellular carcinoma (HCC) are more likely to experience disease progression despite transcatheter arterial chemoembolization (TACE) treatment, and thus would benefit from early switching to other therapeutic regimens. We sought to evaluate a fully automated machine learning algorithm that uses pre-therapeutic quantitative computed tomography (CT) image features and clinical factors to predict HCC response to TACE. Outcome information from 105 patients receiving first-line treatment with TACE was evaluated retrospectively. The primary clinical endpoint was time to progression (TTP) based on follow-up CT radiological criteria (mRECIST). A 14-week cutoff was used to classify patients as TACE-susceptible (TTP ≥14 weeks) or TACE-refractory (TTP <14 weeks). Response to TACE was predicted using a random forest classifier with the Barcelona Clinic Liver Cancer (BCLC) stage and quantitative image features as input as well as the BCLC stage alone as a control. The model's response prediction accuracy rate was 74.2% (95% CI=64%-82%) using a combination of the BCLC stage plus quantitative image features versus 62.9% (95% CI= 52%-72%) using the BCLC stage alone. Shape image features of the tumor and background liver were the dominant features correlated to the TTP as selected by the Boruta method and were used to predict the outcome. This preliminary study demonstrates that quantitative image features obtained prior to therapy can improve the accuracy of predicting response of HCC to TACE. This approach is likely to provide useful information for aiding HCC patient selection for TACE.

Automated Echocardiographic Quantification of Left Ventricular Ejection Fraction Without Volume Measurements Using a Machine Learning Algorithm Mimicking a Human Expert.

Echocardiographic quantification of left ventricular (LV) ejection fraction (EF) relies on either manual or automated identification of endocardial boundaries followed by model-based calculation of end-systolic and end-diastolic LV volumes. Recent developments in artificial intelligence resulted in computer algorithms that allow near automated detection of endocardial boundaries and measurement of LV volumes and function. However, boundary identification is still prone to errors limiting accuracy in certain patients. We hypothesized that a fully automated machine learning algorithm could circumvent border detection and instead would estimate the degree of ventricular contraction, similar to a human expert trained on tens of thousands of images. Machine learning algorithm was developed and trained to automatically estimate LVEF on a database of >50 000 echocardiographic studies, including multiple apical 2- and 4-chamber views (AutoEF, BayLabs). Testing was performed on an independent group of 99 patients, whose automated EF values were compared with reference values obtained by averaging measurements by 3 experts using conventional volume-based technique. Inter-technique agreement was assessed using linear regression and Bland-Altman analysis. Consistency was assessed by mean absolute deviation among automated estimates from different combinations of apical views. Finally, sensitivity and specificity of detecting of EF ≤35% were calculated. These metrics were compared side-by-side against the same reference standard to those obtained from conventional EF measurements by clinical readers. Automated estimation of LVEF was feasible in all 99 patients. AutoEF values showed high consistency (mean absolute deviation =2.9%) and excellent agreement with the reference values: r=0.95, bias=1.0%, limits of agreement =±11.8%, with sensitivity 0.90 and specificity 0.92 for detection of EF ≤35%. This was similar to clinicians' measurements: r=0.94, bias=1.4%, limits of agreement =±13.4%, sensitivity 0.93, specificity 0.87. Machine learning algorithm for volume-independent LVEF estimation is highly feasible and similar in accuracy to conventional volume-based measurements, when compared with reference values provided by an expert panel.

Identification of suspicious invasive placentation based on clinical MRI data using textural features and automated machine learning.

The aim of this study was to investigate whether intraplacental texture features from routine placental MRI can objectively and accurately predict invasive placentation. This retrospective study includes 99 pregnant women with pathologically confirmed placental invasion and 56 pregnant women with simple placenta previa. All participants underwent magnetic resonance imaging after 24 gestational weeks. The placenta was segmented in sagittal images from both turbo spin echo (TSE) and balanced turbo field echo (bTFE) sequences. Textural features were extracted from the both original and Laplacian of Gaussian (LoG)-filtered MRI images. An automated machine learning algorithm was applied to the extracted feature sets to obtain the optimal preprocessing steps, classification algorithm, and corresponding hyper-parameters. A gradient boosting classifier using all textual features from original and LoG-filtered TSE images and bTFE images identified by the automated machine learning algorithm achieved the optimal performance with sensitivity, specificity, accuracy, and area under ROC curve (AUC) of 100%, 88.5%, 95.2%, and 0.98 in the prediction of placental invasion. In addition, textural features that contributed to the prediction of placental invasion differ from the features significantly affected by normal placenta maturation. Quantifying intraplacental heterogeneity using LoG filtration and texture analysis highlights the different heterogeneous appearance caused by abnormal placentation relative to normal maturation. The predictive model derived from automated machine learning yielded good performance, indicating the proposed radiomic analysis pipeline can accurately predict placental invasion and facilitate clinical decision-making for pregnant women with suspicious placental invasion. • The intraplacental texture features have high efficiency in prediction of invasive placentation after 24 gestational weeks. • The features with dominated predictive power did not overlap with the features significantly affected by gestational age.

Predicting Dynamic Fuel Oil Consumption on Ships with Automated Machine Learning

This study demonstrates a method for predicting the dynamic fuel consumption on board ships using automated machine learning algorithms, fed only with data for larger time intervals from 12 hours up to 96 hours. The machine learning algorithm trained on dynamic data from shorter time intervals of the engine features together with longer time interval data for the fuel consumption. To give the operator and ship owner real-time energy efficiency statistics, it is essential to be able to predict the dynamic fuel oil consumption. The conventional approach to getting these data is by installing additional mass flow meters, but these come with added cost and complexity. In this study, we propose a machine learning approach using auto machine learning optimisation, with already available data from the machinery logging system.

Machine learning based automated dynamic quantification of left heart chamber volumes.

Studies have demonstrated the ability of a new automated algorithm for volumetric analysis of 3D echocardiographic (3DE) datasets to provide accurate and reproducible measurements of left ventricular and left atrial (LV, LA) volumes at end-systole and end-diastole. Recently, this methodology was expanded using a machine learning (ML) approach to automatically measure chamber volumes throughout the cardiac cycle, resulting in LV and LA volume-time curves. We aimed to validate ejection and filling parameters obtained from these curves by comparing them to independent well-validated reference techniques. We studied 20 patients referred for cardiac magnetic resonance (CMR) examinations, who underwent 3DE imaging the same day. Volume-time curves were obtained for both LV and LA chambers using the ML algorithm (Philips HeartModel), and independently conventional 3DE volumetric analysis (TomTec), and CMR images (slice-by-slice, frame-by-frame manual tracing). Automatically derived LV and LA volumes and ejection/filling parameters were compared against both reference techniques. Minor manual correction of the automatically detected LV and LA borders was needed in 4/20 and 5/20 cases, respectively. Time required to generate volume-time curves was 35 ± 17 s using ML algorithm, 3.6 ± 0.9 min using conventional 3DE analysis, and 96 ± 14 min using CMR. Volume-time curves obtained by all three techniques were similar in shape and magnitude. In both comparisons, ejection/filling parameters showed no significant inter-technique differences. Bland-Altman analysis confirmed small biases, despite wide limits of agreement. The automated ML algorithm can quickly measure dynamic LV and LA volumes and accurately analyse ejection/filling parameters. Incorporation of this algorithm into the clinical workflow may increase the utilization of 3DE imaging.

Calibrating the HISA temperature: Measuring the temperature of the Riegel–Crutcher cloud

HI self absorption (HISA) clouds are clumps of cold neutral hydrogen (HI) visible in front of warm background gas, which makes them ideal places to study the properties of the cold atomic component of the interstellar medium (ISM). The Riegel-Crutcher (R-C) cloud is the most striking HISA feature in the Galaxy. It is one of the closest HISA clouds to us and is located in the direction of the Galactic Centre, which provides a bright background. High-resolution interferometric measurements have revealed the filamentary structure of this cloud, however it is difficult to accurately determine the temperature and the density of the gas without optical depth measurements. In this paper we present new HI absorption observations with the Australia Telescope Compact Array (ATCA) against 46 continuum sources behind the Riegel-Crutcher cloud to directly measure the optical depth of the cloud. We decompose the complex HI absorption spectra into Gaussian components using an automated machine learning algorithm. We find 300 Gaussian components, from which 67 are associated with the R-C cloud (0 < v_{LSR} < 10 kms, FWHM < 10 kms). Combining the new HI absorption data with HI emission data from previous surveys we calculate the spin temperature and find it to be between 20 and 80 K. Our measurements uncover a temperature gradient across the cloud with spin temperatures decreasing towards positive Galactic latitudes. We also find three new OH absorption lines associated with the cloud, which support the presence of molecular gas.

Development and Evaluation of an Automated Machine Learning Algorithm for In-Hospital Mortality Risk Adjustment Among Critical Care Patients.

Risk adjustment algorithms for ICU mortality are necessary for measuring and improving ICU performance. Existing risk adjustment algorithms are not widely adopted. Key barriers to adoption include licensing and implementation costs as well as labor costs associated with human-intensive data collection. Widespread adoption of electronic health records makes automated risk adjustment feasible. Using modern machine learning methods and open source tools, we developed and evaluated a retrospective risk adjustment algorithm for in-hospital mortality among ICU patients. The Risk of Inpatient Death score can be fully automated and is reliant upon data elements that are generated in the course of usual hospital processes. One hundred thirty-one ICUs in 53 hospitals operated by Tenet Healthcare. A cohort of 237,173 ICU patients discharged between January 2014 and December 2016. The data were randomly split into training (36 hospitals), and validation (17 hospitals) data sets. Feature selection and model training were carried out using the training set while the discrimination, calibration, and accuracy of the model were assessed in the validation data set. Model discrimination was evaluated based on the area under receiver operating characteristic curve; accuracy and calibration were assessed via adjusted Brier scores and visual analysis of calibration curves. Seventeen features, including a mix of clinical and administrative data elements, were retained in the final model. The Risk of Inpatient Death score demonstrated excellent discrimination (area under receiver operating characteristic curve = 0.94) and calibration (adjusted Brier score = 52.8%) in the validation dataset; these results compare favorably to the published performance statistics for the most commonly used mortality risk adjustment algorithms. Low adoption of ICU mortality risk adjustment algorithms impedes progress toward increasing the value of the healthcare delivered in ICUs. The Risk of Inpatient Death score has many attractive attributes that address the key barriers to adoption of ICU risk adjustment algorithms and performs comparably to existing human-intensive algorithms. Automated risk adjustment algorithms have the potential to obviate known barriers to adoption such as cost-prohibitive licensing fees and significant direct labor costs. Further evaluation is needed to ensure that the level of performance observed in this study could be achieved at independent sites.

Automatic assessment of mental health using phone metadata

ABSTRACTAn accurate and timely assessment of mental health of individuals is an important building block toward healthy well‐functioning societies. Traditional methods for assessing mental health and wellbeing are surveys administered by health‐care professionals. Such surveys are labor intensive, costly, and typically intermittent. With the growth in mobile sensing, health informatics, and data science, this work argues a case for using mobile phone meta‐data and data analytics to automatically infer an individual's mental health. Specifically, based on a clinically utilized health measure as ground truth (MHI‐5) and ten‐week phone usage metadata corpus for 45 participants we report that automated machine learning algorithms utilizing phone based features can achieve reasonable accuracy (around 80%) at automatically classifying the level of mental health of an individual. Results could pave way for cheaper, automated, health assessments with timely escalations to mental health professionals and interventions when required.

Ultrasound-based gestational-age estimation in late pregnancy.

ABSTRACTObjectiveAccurate gestational‐age (GA) estimation, preferably by ultrasound measurement of fetal crown–rump length before 14 weeks' gestation, is an important component of high‐quality antenatal care. The objective of this study was to determine how GA can best be estimated by fetal ultrasound for women who present for the first time late in pregnancy with uncertain or unknown menstrual dates.MethodsINTERGROWTH‐21st was a large, prospective, multicenter, population‐based project performed in eight geographically defined urban populations. One of its principal components, the Fetal Growth Longitudinal Study, aimed to develop international fetal growth standards. Each participant had their certain menstrual dates confirmed by first‐trimester ultrasound examination. Fetal head circumference (HC), biparietal diameter (BPD), occipitofrontal diameter (OFD), abdominal circumference (AC) and femur length (FL) were measured every 5 weeks from 14 weeks' gestation until delivery. For each participant, a single, randomly selected ultrasound examination was used to explore all candidate biometric variables and permutations to build models to predict GA. Regression equations were ranked based upon minimization of the mean prediction error, goodness of fit and model complexity. An automated machine learning algorithm, the Genetic Algorithm, was adapted to evaluate > 64 000 potential polynomial equations as predictors.ResultsOf the 4607 eligible women, 4321 (94%) had a pregnancy without major complications and delivered a live singleton without congenital malformations. After other exclusions (missing measurements in GA window and outliers), the final sample comprised 4229 women. Two skeletal measures, HC and FL, produced the best GA prediction, given by the equation loge(GA) = 0.03243 × (loge(HC))2 + 0.001644 × FL × loge(HC) + 3.813. When FL was not available, the best equation based on HC alone was loge(GA) = 0.05970 × (loge(HC))2 + 0.000000006409 × (HC)3 + 3.3258. The estimated uncertainty of GA prediction (half width 95% interval) was 6–7 days at 14 weeks' gestation, 12–14 days at 26 weeks' gestation and > 14 days in the third trimester. The addition of FL to the HC model led to improved prediction intervals compared with using HC alone, but no further improvement in prediction was afforded by adding AC, BPD or OFD. Equations that included other measurements (BPD, OFD and AC) did not perform better.ConclusionsAmong women initiating antenatal care late in pregnancy, a single set of ultrasound measurements combining HC and FL in the second trimester can be used to estimate GA with reasonable accuracy. We recommend this tool for underserved populations but considerable efforts should be implemented to improve early initiation of antenatal care worldwide. © 2016 Authors. Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of International Society of Ultrasound in Obstetrics and Gynecology.

Data-driven estimation of blood pressure using photoplethysmographic signals.

Noninvasive measurement of blood pressure by optical methods receives considerable interest, but the complexity of the measurement and the difficulty of adjusting parameters restrict applications. We develop a method for estimating the systolic and diastolic blood pressure using a single-point optical recording of a photoplethysmographic (PPG) signal. The estimation is data-driven, we use automated machine learning algorithms instead of mathematical models. Combining supervised learning with a discrete wavelet transform, the method is insensitive to minor irregularities in the PPG waveform, hence both pulse oximeters and smartphone cameras can record the signal. We evaluate the accuracy of the estimation on 78 samples from 65 subjects (40 male, 25 female, age 29±7) with no history of cardiovascular disease. The estimate for systolic blood pressure has a mean error 4.9±4.9 mm Hg, and 4.3±3.7 mm Hg for diastolic blood pressure when using the oximeter-obtained PPG. The same values are 5.1±4.3 mm Hg and 4.6±4.3 mm Hg when using the phone-obtained PPG, comparing with A&D UA-767PBT result as gold standard. The simplicity of the method encourages ambulatory measurement, and given the ease of sharing the measured data, we expect a shift to data-oriented approaches deriving insight from ubiquitous mobile devices that will yield more accurate machine learning models in monitoring blood pressure.

Abstract 122: Use of Machine Learning to Accurately Predict Adverse Events in Patients With Peripheral Artery Disease Using Electronic Health Record Data

Introduction: Patients with peripheral artery disease (PAD) are at high risk of major adverse cardiac and cerebrovascular events (MACCE). However, no currently available risk scores accurately delineate which patients are most likely to sustain an event, creating a missed opportunity for more aggressive risk factor management. We set out to develop a novel predictive model - based on automated machine learning algorithms using electronic health record (EHR) data - with the aim of identifying which PAD patients are most likely to have an adverse outcome during follow-up. Methods: Data were derived from patients with a diagnosis of PAD at our institution. Novel machine-learning algorithms including random forest and penalized regression predictive models were developed using structured and unstructured data that including lab values, diagnosis codes, medications, and clinical notes. Patients were matched for total follow-up time to remove length of patient records as a biasing factor in our predictive models. Results: After matching for length of follow-up, 3,807 patients were included in our models. A total of 1,269 patients had a MACCE event after PAD diagnosis. The median time to MACCE was 2.8 years after PAD diagnosis. Utilizing 1,492 different variables extracted from the EHR, our best predictive model was able to very accurately predict which patients would go on to have a MACCE event after diagnosis of PAD with an AUC of 0.98, with a sensitivity, specificity and positive predictive value of 0.90, 0.96, and 0.93, respectively. Conclusions: Hypothesis-free, machine-learning algorithms using freely available data in the EHR can accurately predict which PAD patients are most likely to go on to develop future MACCE. While these findings require validation in an independent data set, there is hope that these informatics approaches can be applied to the medical record in an automated fashion to risk stratify patients with vascular disease and identify those who might benefit from more aggressive disease management in real-time.

Outlier Detection in OpenStreetMap Data using the Random Forest Algorithm

GIScience 2016 Short Paper Proceedings Outlier Detection in OpenStreetMap Data using the Random Forest Algorithm Richard Wen, Claus Rinner Department of Geography and Environmental Studies, Ryerson University 350 Victoria St., Toronto, Ontario, M5B 2K3, Canada Email: {rwen, crinner}@ryerson.ca Abstract OpenStreetMap (OSM) data consist of digitized geographic objects with semantic tags assigned by the volunteer contributors. The tags describe the geographic objects in a way that is understandable by both humans and computers. The variability in contributor behaviour creates reliability concerns for the tagging quality of OSM data. The detection of irregular contributions may improve OSM data quality and editing tools. This research applies the random forest algorithm on geospatial variables in order to detect outliers without ground-truth reference data to direct human inspection. An application to OSM data for Toronto, Ontario, was effective in revealing abnormal amenity tagging of school and hospital objects. 1. Introduction OpenStreetMap (OSM) is an online platform enabling registered volunteers to contribute geospatial data by digitizing point-, line-, or polygon-shaped geographic objects and annotating them with tags referring to common feature classes such as roads and restaurants (Haklay 2008). OSM tags are semantically structured as key-value pairs, where the key refers to a broad class of geographic objects and the value details the specific geographic object being tagged (Ballatore et al. 2013). Examples of tags are amenity=school, highway=residential, and building=house. The open and flexible nature of OSM tagging leads to varying contribution behaviour by different communities (Mooney et al. 2010). The varying contribution behaviour creates concerns about the quality of OSM data and the community standards of OSM tagging. Quality control and corrections rely heavily on human interaction, which raises additional questions on the reliability of OSM data. Finally, the experience of the volunteer contributor has an effect on the tagging quality of each geographic object as experienced contributors are more familiar with the tagging norms of the area being edited. Although OSM is an effective and efficient platform for generating masses of geospatial data, it is plagued by reliability, quality, and completeness issues. The aim of this paper is to examine the ability of an automated machine learning algorithm, the random forest algorithm, to support manual human inspection and minimize bias in OSM data editing. The use of an automated algorithm improves the detection of abnormal tagging behaviour, avoids the bias of human judgement, and reduces the time required to search through masses of tagged geographic objects. A combination of human knowledge and experience with the logical accuracy of machines could improve OSM tagging quality and standards, and enable the development of advanced editing tools. 2. Data and Methods

Supervised prediction of drug-induced nephrotoxicity based on interleukin-6 and -8 expression levels.

BackgroundDrug-induced nephrotoxicity causes acute kidney injury and chronic kidney diseases, and is a major reason for late-stage failures in the clinical trials of new drugs. Therefore, early, pre-clinical prediction of nephrotoxicity could help to prioritize drug candidates for further evaluations, and increase the success rates of clinical trials. Recently, an in vitro model for predicting renal-proximal-tubular-cell (PTC) toxicity based on the expression levels of two inflammatory markers, interleukin (IL)-6 and -8, has been described. However, this and other existing models usually use linear and manually determined thresholds to predict nephrotoxicity. Automated machine learning algorithms may improve these models, and produce more accurate and unbiased predictions.ResultsHere, we report a systematic comparison of the performances of four supervised classifiers, namely random forest, support vector machine, k-nearest-neighbor and naive Bayes classifiers, in predicting PTC toxicity based on IL-6 and -8 expression levels. Using a dataset of human primary PTCs treated with 41 well-characterized compounds that are toxic or not toxic to PTC, we found that random forest classifiers have the highest cross-validated classification performance (mean balanced accuracy = 87.8%, sensitivity = 89.4%, and specificity = 85.9%). Furthermore, we also found that IL-8 is more predictive than IL-6, but a combination of both markers gives higher classification accuracy. Finally, we also show that random forest classifiers trained automatically on the whole dataset have higher mean balanced accuracy than a previous threshold-based classifier constructed for the same dataset (99.3% vs. 80.7%).ConclusionsOur results suggest that a random forest classifier can be used to automatically predict drug-induced PTC toxicity based on the expression levels of IL-6 and -8.

Decision tree for selecting retaining wall systems based on logistic regression analysis

Machine learning techniques generally require thousands of cases to derive a reliable conclusion, but such a large number of excavation cases are very difficult to acquire in the construction domain. There have been efforts to develop retaining wall selection systems using machine learning techniques but based only on a couple of hundred cases of excavation work. The resultant rules were inconsistent and unreliable. This paper proposes an improved decision tree for selecting retaining wall systems. After retaining wall systems were divided into three components, i.e., the retaining wall, the lateral support, and optional grouting, a series of logistic regression analyses, analysis of variance (ANOVA), and chi-square tests were used to derive the variables and a decision tree for selecting retaining wall systems. The prediction accuracy rates for the retaining walls, lateral supports, and grouting were 82.6%, 80.4%, and 76.9%, respectively. These values were higher than the prediction accuracy rate (58.7%) of the decision tree built by an automated machine learning algorithm, Classification and Regression Trees (CART), with the same data set.

Automated 3D mapping of baseline and 12-month associations between three verbal memory measures and hippocampal atrophy in 490 ADNI subjects

We used a previously validated automated machine learning algorithm based on adaptive boosting to segment the hippocampi in baseline and 12-month follow-up 3D T1-weighted brain MRIs of 150 cognitively normal elderly (NC), 245 mild cognitive impairment (MCI) and 97 Dementia of the Alzheimer’s type (DAT) ADNI subjects. Using the radial distance mapping technique, we examined the hippocampal correlates of delayed recall performance on three well-established verbal memory tests—ADAScog delayed recall (ADAScog-DR), the Rey Auditory Verbal Learning Test -DR (AVLT-DR) and Wechsler Logical Memory II-DR (LM II-DR). We observed no significant correlations between delayed recall performance and hippocampal radial distance on any of the three verbal memory measures in NC. All three measures were associated with hippocampal volumes and radial distance in the full sample and in the MCI group at baseline and at follow-up. In DAT we observed stronger left-sided associations between hippocampal radial distance, LM II-DR and ADAScog-DR both at baseline and at follow-up. The strongest linkage between memory performance and hippocampal atrophy in the MCI sample was observed with the most challenging verbal memory test—the AVLT-DR, as opposed to the DAT sample where the least challenging test the ADAScog-DR showed strongest associations with the hippocampal structure. After controlling for baseline hippocampal atrophy, memory performance showed regionally specific associations with hippocampal radial distance in predominantly CA1 but also in subicular distribution.

IC-P3-223: Automated longitudinal 3D mapping of hippocampal ADASCPG delayed recall effects in 293 normal elderly, MCI and AD subjects

Background: The Alzheimer’s Disease Assessment Scale Cognitive Subscale (ADAScog) is a primary outcome measure in Alzheimer’s disease (AD) and mild cognitive impairment (MCI) clinical trials. Hippocampal atrophy has recently been proposed as a sensitive biomarker in AD and MCI. Methods: We applied a previously validated automated machine learning algorithm based on adaptive boosting (Morra et al., 2008) to segment the hippocampi in baseline and 12-month follow-up 3D T1-weighted brain MRIs of 109 cognitively normal elderly (NC), 133 MCI and 51 AD subjects from the ADNI study. We then applied the radial distance atrophy mapping approach to examine in 3D the structural hippocampal correlates of ADAScog delayed recall (ADAScog-DR). 3D statistical maps were corrected for multiple comparisons using permutation testing thresholded at p 0.01. Results: The groups were well balanced with regard to age, race and education. The MCI group had significantly more men (71%) relative to the NC (52%) and the AD group (51%). The MCI and AD groups showed a significant decline in ADAScog-DR over 12 months (MCI 0.42, p 0.012; AD 0.49, p 0.005). At baseline ADAScog-DR showed significant correlations with atrophy of CA1 and subicular hippocampal subregions in MCI (r 0.2-0.3, left pcorrected 0.03; right pcorrected 0.04) and the right mid/posterior CA1, CA2-3 and subiculum in AD (r 0.35-0.55, pcorrected 0.0004). At followup, ADAScog-DR correlated with atrophy of the left more then right CA1 and subiculum in MCI (r 0.2-0.3, pcorrected 0.003) and the right anterior CA1 and subiculum in AD (r 0.25-0.45; pcorrected 0.017). Correlations between baseline ADAScog-DR and posterior right hippocampal atrophy were observed at trend level in NC (r 0.2-0.4; pcorrected 0.07). Correlations with the global ADAScog ratings (ADAScog11 and ADAScog-modified) were generally less significant than those observed with ADAScog-DR. Conclusions: We demonstrated regionally specific correlations between ADAScog-DR and hippocampal atrophy in MCI and AD. The MCI 3D maps revealed correlations with hippocampal subregions known to be affected early in MCI. Our high throughput automated segmentation approach is a promising technique for analyzing large epidemiological and clinical trial datasets. IC-P3-224 REGIONAL VARIATIONS IN AGE DEPENDENCE OF P-GP MEDIATED TRANSPORT ACROSS THE BLOOD-BRAIN BARRIER MEASURED USING (R)[C]VERAPAMIL AND POSITRON EMISSION TOMOGRAPHY

Mathematical Morphology: Star/Galaxy Differentiation &amp; Galaxy Morphology Classification

We present an application of Mathematical Morphology (MM) for the classification of astronomical objects, both for star/galaxy differentiation and galaxy morphology classification. We demonstrate that, for CCD images, 99.3 ± 3.8% of galaxies can be separated from stars using MM, with 19.4 ± 7.9% of the stars being misclassified. We demonstrate that, for photographic plate images, the number of galaxies correctly separated from the stars can be increased using our MM diffraction spike tool, which allows 51.0 ± 6.0% of the high-brightness galaxies that are inseparable in current techniques to be correctly classified, with only 1.4 ± 0.5% of the high-brightness stars contaminating the population. We demonstrate that elliptical (E) and late-type spiral (Sc-Sd) galaxies can be classified using MM with an accuracy of 91.4 ± 7.8%. It is a method involving fewer ‘free parameters’ than current techniques, especially automated machine learning algorithms. The limitation of MM galaxy morphology classification based on seeing and distance is also presented. We examine various star/galaxy differentiation and galaxy morphology classification techniques commonly used today, and show that our MM techniques compare very favourably.