Use of machine learning to identify characteristics associated with severe hypoglycemia in older adults with type 1 diabetes: a post-hoc analysis of a case-control study

INTRODUCTION

Severe hypoglycemia (SH) in older adults (OAs) with type 1 diabetes is associated with profound morbidity and mortality, yet its etiology can be complex and multifactorial. Enhanced tools to identify OAs who are at high risk for SH are needed. This study used machine learning to identify characteristics that distinguish those with and without recent SH, selecting from a range of demographic and clinical, behavioral and lifestyle, and neurocognitive characteristics, along with continuous glucose monitoring (CGM) measures.

RESEARCH DESIGN AND METHODS

Data from a case-control study involving OAs recruited from the T1D Exchange Clinical Network were analyzed. The random forest machine learning algorithm was used to elucidate the characteristics associated with case versus control status and their relative importance. Models with successively rich characteristic sets were examined to systematically incorporate each domain of possible risk characteristics.

RESULTS

Data from 191 OAs with type 1 diabetes (47.1% female, 92.1% non-Hispanic white) were analyzed. Across models, hypoglycemia unawareness was the top characteristic associated with SH history. For the model with the richest input data, the most important characteristics, in descending order, were hypoglycemia unawareness, hypoglycemia fear, coefficient of variation from CGM, % time blood glucose below 70 mg/dL, and trail making test B score.

CONCLUSIONS

Machine learning may augment risk stratification for OAs by identifying key characteristics associated with SH. Prospective studies are needed to identify the predictive performance of these risk characteristics.

Real-time calculation of a limiting form of the Renyi entropy applied to detection of subtle changes in scattering architecture

Previously a new method for ultrasound signal characterization using entropy H(f) was reported, and it was demonstrated that in certain settings, further improvements in signal characterization could be obtained by generalizing to Renyi entropy-based signal characterization I(f)(r) with values of r near 2 (specifically r=1.99) [M. S. Hughes et al., J. Acoust. Soc. Am. 125, 3141-3145 (2009)]. It was speculated that further improvements in sensitivity might be realized at the limit r-->2. At that time, such investigation was not feasible due to excessive computational time required to calculate I(f)(r) near this limit. In this paper, an asymptotic expression for the limiting behavior of I(f)(r) as r-->2 is derived and used to present results analogous to those obtained with I(f)(1.99). Moreover, the limiting form I(f,infinity) is computable directly from the experimentally measured waveform f(t) by an algorithm that is suitable for real-time calculation and implementation.

Fast wavelet estimation of weak biosignals

Wavelet-based signal processing has become commonplace in the signal processing community over the past decade and wavelet-based software tools and integrated circuits are now commercially available. One of the most important applications of wavelets is in removal of noise from signals, called denoising, accomplished by thresholding wavelet coefficients in order to separate signal from noise. Substantial work in this area was summarized by Donoho and colleagues at Stanford University, who developed a variety of algorithms for conventional denoising. However, conventional denoising fails for signals with low signal-to-noise ratio (SNR). Electrical signals acquired from the human body, called biosignals, commonly have below 0 dB SNR. Synchronous linear averaging of a large number of acquired data frames is universally used to increase the SNR of weak biosignals. A novel wavelet-based estimator is presented for fast estimation of such signals. The new estimation algorithm provides a faster rate of convergence to the underlying signal than linear averaging. The algorithm is implemented for processing of auditory brainstem response (ABR) and of auditory middle latency response (AMLR) signals. Experimental results with both simulated data and human subjects demonstrate that the novel wavelet estimator achieves superior performance to that of linear averaging.

A comparison of Monte Carlo dose calculation denoising techniques

Recent studies have demonstrated that Monte Carlo (MC) denoising techniques can reduce MC radiotherapy dose computation time significantly by preferentially eliminating statistical fluctuations ('noise') through smoothing. In this study, we compare new and previously published approaches to MC denoising, including 3D wavelet threshold denoising with sub-band adaptive thresholding, content adaptive mean-median-hybrid (CAMH) filtering, locally adaptive Savitzky-Golay curve-fitting (LASG), anisotropic diffusion (AD) and an iterative reduction of noise (IRON) method formulated as an optimization problem. Several challenging phantom and computed-tomography-based MC dose distributions with varying levels of noise formed the test set. Denoising effectiveness was measured in three ways: by improvements in the mean-square-error (MSE) with respect to a reference (low noise) dose distribution; by the maximum difference from the reference distribution and by the 'Van Dyk' pass/fail criteria of either adequate agreement with the reference image in low-gradient regions (within 2% in our case) or, in high-gradient regions, a distance-to-agreement-within-2% of less than 2 mm. Results varied significantly based on the dose test case: greater reductions in MSE were observed for the relatively smoother phantom-based dose distribution (up to a factor of 16 for the LASG algorithm); smaller reductions were seen for an intensity modulated radiation therapy (IMRT) head and neck case (typically, factors of 2-4). Although several algorithms reduced statistical noise for all test geometries, the LASG method had the best MSE reduction for three of the four test geometries, and performed the best for the Van Dyk criteria. However, the wavelet thresholding method performed better for the head and neck IMRT geometry and also decreased the maximum error more effectively than LASG. In almost all cases, the evaluated methods provided acceleration of MC results towards statistically more accurate results.

Accelerating Monte Carlo simulations of radiation therapy dose distributions using wavelet threshold de-noising

The Monte Carlo dose calculation method works by simulating individual energetic photons or electrons as they traverse a digital representation of the patient anatomy. However, Monte Carlo results fluctuate until a large number of particles are simulated. We propose wavelet threshold de-noising as a postprocessing step to accelerate convergence of Monte Carlo dose calculations. A sampled rough function (such as Monte Carlo noise) gives wavelet transform coefficients which are more nearly equal in amplitude than those of a sampled smooth function. Wavelet hard-threshold de-noising sets to zero those wavelet coefficients which fall below a threshold; the image is then reconstructed. We implemented the computationally efficient 9,7-biorthogonal filters in the C language. Transform results were averaged over transform origin selections to reduce artifacts. A method for selecting best threshold values is described. The algorithm requires about 336 floating point arithmetic operations per dose grid point. We applied wavelet threshold de-noising to two two-dimensional dose distributions: a dose distribution generated by 10 MeV electrons incident on a water phantom with a step-heterogeneity, and a slice from a lung heterogeneity phantom. Dose distributions were simulated using the Integrated Tiger Series Monte Carlo code. We studied threshold selection, resulting dose image smoothness, and resulting dose image accuracy as a function of the number of source particles. For both phantoms, with a suitable value of the threshold parameter, voxel-to-voxel noise was suppressed with little introduction of bias. The roughness of wavelet de-noised dose distributions (according to a Laplacian metric) was nearly independent of the number of source electrons, though the accuracy of the de-noised dose image improved with increasing numbers of source electrons. We conclude that wavelet shrinkage de-noising is a promising method for effectively accelerating Monte Carlo dose calculations by factors of 2 or more.

Wavelets, adapted waveforms and de-noising

This is a short summary of a talk given at the Frontier Science in EEG Symposium, Continuous Waveform Analysis, held on 9 October 1993 in New Orleans. We describe some new libraries of waveforms well-adapted to various numerical analysis and signal processing tasks. The main point is that by expanding a signal in a library of waveforms which are well-localized in both time and frequency, one can achieve both understanding of structure and efficiency in computation. We briefly cover the properties of the new "wavelet packet" and "localized trigonometric" libraries. The main focus will be applications of such libraries to the analysis of complicated transient signals: a feature extraction and data compression algorithm for speech signals which uses best-adapted time and frequency decompositions, and an adapted waveform analysis algorithm for removing fish noises from hydrophone recordings. These signals share many of the same properties as EEG traces, but with distinct features that are easier to characterize and detect.