Hiram A. Firpi

Baseline MRI Predictors of Conversion from MCI to Probable AD in the ADNI Cohort

The Alzheimer’s Disease Neuroimaging Initiative (ADNI) is a multi-center study assessing neuroimaging in diagnosis and longitudinal monitoring. Amnestic Mild Cognitive Impairment (MCI) often represents a prodromal form of dementia, conferring a 10-15% annual risk of converting to probable AD. We analyzed baseline 1.5T MRI scans in 693 participants from the ADNI cohort divided into four groups by baseline diagnosis and one year MCI to probable AD conversion status to identify neuroimaging phenotypes associated with MCI and AD and potential predictive markers of imminent conversion. MP-RAGE scans were analyzed using publicly available voxel-based morphometry (VBM) and automated parcellation methods. Measures included global and hippocampal grey matter (GM) density, hippocampal and amygdalar volumes, and cortical thickness values from entorhinal cortex and other temporal and parietal lobe regions. The overall pattern of structural MRI changes in MCI (n=339) and AD (n=148) compared to healthy controls (HC, n=206) was similar to prior findings in smaller samples. MCI-Converters (n=62) demonstrated a very similar pattern of atrophic changes to the AD group up to a year before meeting clinical criteria for AD. Finally, a comparison of effect sizes for contrasts between the MCI-Converters and MCI-Stable (n=277) groups on MRI metrics indicated that degree of neurodegeneration of medial temporal structures was the best antecedent MRI marker of imminent conversion, with decreased hippocampal volume (left > right) being the most robust. Validation of imaging biomarkers is important as they can help enrich clinical trials of disease modifying agents by identifying individuals at highest risk for progression to AD.

Swarmed feature selection

Feature selection is an important part of pattern recognition, helping to overcome the curse of dimensionality problem with classifiers, among other systems. In this work, we introduce a feature selection method using particle swarm optimization. Experiments using data of others and hyperspectral remote sensed data are used to measure the performance of the algorithm. Its comparison with a genetic algorithm is also shown.

Comparison of Manual and Automated Determination of Hippocampal Volumes in MCI and Early AD

MRI-based hippocampal volume analysis has been extensively employed given its potential as a biomarker for brain disorders such as Alzheimer’s disease (AD), and accurate and efficient determination of hippocampal volumes from brain images is still a challenging issue. We compared an automated method, FreeSurfer (V4), with a published manual protocol for the determination of hippocampal volumes from T1-weighted MRI scans. Our study included MRI data from 125 older adult subjects: healthy controls with no significant cognitive complaints or deficits (HC, n = 38), euthymic individuals with cognitive complaints (CC, n = 39) but intact neuropsychological performance, and patients with amnestic mild cognitive impairment (MCI, n = 37) or a clinical diagnosis of probable AD (AD, n = 11). Pearson correlations and intraclass correlation coefficients (ICCs) were calculated to evaluate the relationship between results of the manual tracing and FreeSurfer methods and to estimate their agreement. Results indicated that these two methods derived highly correlated results with strong agreement. After controlling for the age, sex and intracranial volume in statistical group analysis, both the manual tracing and FreeSurfer methods yield similar patterns: both the MCI group and the AD group showed hippocampal volume reduction compared to both the HC group and the CC group, and the HC and CC groups did not differ. These comparisons suggest that FreeSurfer has the potential to be used in automated determination of hippocampal volumes for large-scale MCI/AD-related MRI studies, where manual methods are inefficient or not feasible.

Epileptic seizure detection by means of genetically programmed artificial features

In this paper, we describe a general-purpose, systematic algorithm, consisting of a genetic programming module and a k-nearest neighbor classifier to automatically create artificial features-features that are computer-crafted and may not have a known physical meaning-directly from the reconstructed state-space trajectories of the EEG signals that reveal patterns indicative of epileptic seizure onset. The algorithm was evaluated in three patients and validation experiments were carried out using 267.6 hours of EEG recordings. The results with the artificial features compare favorably with previous benchmark work that used a handcrafted feature.

Parametric surface modeling and registration for comparison of manual and automated segmentation of the hippocampus

Accurate and efficient segmentation of the hippocampus from brain images is a challenging issue. Although experienced anatomic tracers can be reliable, manual segmentation is a time consuming process and may not be feasible for large‐scale neuroimaging studies. In this article, we compare an automated method, FreeSurfer (V4), with a published manual protocol on the determination of hippocampal boundaries from magnetic resonance imaging scans, using data from an existing mild cognitive impairment/Alzheimers disease cohort. To perform the comparison, we develop an enhanced spherical harmonic processing framework to model and register these hippocampal traces. The framework treats the two hippocampi as a single geometric configuration and extracts the positional, orientation, and shape variables in a multiobject setting. We apply this framework to register manual tracing and FreeSurfer results together and the two methods show stronger agreement on position and orientation than shape measures. Work is in progress to examine a refined FreeSurfer segmentation strategy and an improved agreement on shape features is expected.

Genetic programming of conventional features to detect seizure precursors

This paper presents an application of genetic programming (GP) to optimally select and fuse conventional features (C-features) for the detection of epileptic waveforms within intracranial electroencephalogram (IEEG) recordings that precede seizures, known as seizure-precursors. Evidence suggests that seizure-precursors may localize regions important to seizure generation on the IEEG and epilepsy treatment. However, current methods to detect epileptic precursors lack a sound approach to automatically select and combine C-features that best distinguish epileptic events from background, relying on visual review predominantly. This work suggests GP as an optimal alternative to create a single feature after evaluating the performance of a binary detector that uses: 1) genetically programmed features; 2) features selected via GP; 3) forward sequentially selected features; and 4) visually selected features. Results demonstrate that a detector with a genetically programmed feature outperforms the other three approaches, achieving over 78.5% positive predictive value, 83.5% sensitivity, and 93% specificity at the 95% level of confidence.

Genetic Programming Artificial Features with Applications to Epileptic Seizure Prediction

In this paper, we propose a general-purpose, systematic algorithm, consisting of a genetic programming module and a k-nearest neighbor classifier to automatically create artificial features (i.e., features that are computer crafted and may not have a known physical meaning) directly from the reconstructed state-space trajectory of the EEG signals that reveal patterns predictive of epileptic seizures. The algorithm was evaluated in three different patients, with prediction defined over a horizon of 5 minutes before unequivocal electrographic onset. Experiments are carried out using 20 baseline epochs (non-seizures) and 18 preictal epochs (pre-seizures). Results show that just two seizures were missed while a perfect classification on the baseline epochs was achieved, yielding a 0.0 false positive per hour

On prediction of epileptic seizures by computing multiple genetic programming artificial features

In this paper, we present a general-purpose, systematic algorithm, consisting of a genetic programming module and a k-nearest neighbor classifier, to automatically create multiple artificial features (i.e., features that are computer-crafted and may not have a known physical meaning) directly from EEG signals, in a process that reveals patterns predictive of epileptic seizures. The algorithm was evaluated in three patients, with prediction defined over a horizon that varies between 1 and 5 minutes before unequivocal electrographic onset of seizure. For one patient, a perfect classification was achieved. For the other two patients, high classification accuracy was reached, predicting three seizures out of four for one, and eleven seizures out of fifteen for the other. For the latter, also, only one normal (non-seizure) signal was misclassified. These results compare favorably with other prediction approaches for patients from the same population.

Designing templates for cellular neural networks using particle swarm optimization

Designing or learning of templates for cellular neural networks constitutes one of the crucial research problems of this paradigm. In this work, we present the use of a particle swarm optimizer, a global search algorithm, to design a template set for a CNN. A brief overview of the algorithms and methods is given. Design of popular templates is performed using the search algorithm described.

Prediction of atrial fibrillation following cardiac surgery using rough set derived rules

Atrial fibrillation (AF) and flutter are common following cardiac surgery, increasing costs and morbidity. Cardiologists need a method to discern those patients who are at high risk for this arrhythmia in order to attempt to treat them by either pharmacologic or non-pharmacologic means. We performed a retrospective analysis of 377 CABG patients, of which 94 developed AF post-operatively. Feature selection and AF occurrence prediction was performed using a multivariate regression model, and two rough set derived rule classifiers. The rough set derived feature subset performed best with an accuracy of 87%, a sensitivity of 58.5%, and a specificity of 96.5%. This shows the importance of testing feature subsets, thereby discouraging the practice of simply combining the best individual predictors. The utility of rough set theory in prediction of cardiac arrhythmia is also validated