Seth R. Flaxman

European Union Regulations on Algorithmic Decision-Making and a “Right to Explanation”

We summarize the potential impact that the European Union’s new General Data Protection Regulation will have on the routine use of machine learning algorithms. Slated to take effect as law across the EU in 2018, it will restrict automated individual decision-making (that is, algorithms that make decisions based on user-level predictors) which “significantly affect” users. The law will also effectively create a “right to explanation,” whereby a user can ask for an explanation of an algorithmic decision that was made about them. We argue that while this law will pose large challenges for industry, it highlights opportunities for computer scientists to take the lead in designing algorithms and evaluation frameworks which avoid discrimination and enable explanation.

Global Vision Impairment and Blindness Due to Uncorrected Refractive Error, 1990-2010.

&NA; The purpose of this systematic review was to estimate worldwide the number of people with moderate and severe visual impairment (MSVI; presenting visual acuity <6/18, ≥3/60) or blindness (presenting visual acuity <3/60) due to uncorrected refractive error (URE), to estimate trends in prevalence from 1990 to 2010, and to analyze regional differences. The review focuses on uncorrected refractive error which is now the most common cause of avoidable visual impairment globally. &NA; The systematic review of 14,908 relevant manuscripts from 1990 to 2010 using Medline, Embase, and WHOLIS yielded 243 high-quality, population-based cross-sectional studies which informed a meta-analysis of trends by region. The results showed that in 2010, 6.8 million (95% confidence interval [CI]: 4.7–8.8 million) people were blind (7.9% increase from 1990) and 101.2 million (95% CI: 87.88–125.5 million) vision impaired due to URE (15% increase since 1990), while the global population increased by 30% (1990–2010). The all-age age-standardized prevalence of URE blindness decreased 33% from 0.2% (95% CI: 0.1–0.2%) in 1990 to 0.1% (95% CI: 0.1–0.1%) in 2010, whereas the prevalence of URE MSVI decreased 25% from 2.1% (95% CI: 1.6–2.4%) in 1990 to 1.5% (95% CI: 1.3–1.9%) in 2010. In 2010, URE contributed 20.9% (95% CI: 15.2–25.9%) of all blindness and 52.9% (95% CI: 47.2–57.3%) of all MSVI worldwide. The contribution of URE to all MSVI ranged from 44.2 to 48.1% in all regions except in South Asia which was at 65.4% (95% CI: 62–72%). &NA; We conclude that in 2010, uncorrected refractive error continues as the leading cause of vision impairment and the second leading cause of blindness worldwide, affecting a total of 108 million people or 1 in 90 persons.

Improved prediction accuracy for disease risk mapping using Gaussian process stacked generalization

Maps of infectious disease—charting spatial variations in the force of infection, degree of endemicity and the burden on human health—provide an essential evidence base to support planning towards global health targets. Contemporary disease mapping efforts have embraced statistical modelling approaches to properly acknowledge uncertainties in both the available measurements and their spatial interpolation. The most common such approach is Gaussian process regression, a mathematical framework composed of two components: a mean function harnessing the predictive power of multiple independent variables, and a covariance function yielding spatio-temporal shrinkage against residual variation from the mean. Though many techniques have been developed to improve the flexibility and fitting of the covariance function, models for the mean function have typically been restricted to simple linear terms. For infectious diseases, known to be driven by complex interactions between environmental and socio-economic factors, improved modelling of the mean function can greatly boost predictive power. Here, we present an ensemble approach based on stacked generalization that allows for multiple nonlinear algorithmic mean functions to be jointly embedded within the Gaussian process framework. We apply this method to mapping Plasmodium falciparum prevalence data in sub-Saharan Africa and show that the generalized ensemble approach markedly outperforms any individual method.

Prevalence and causes of vision loss in high-income countries and in Eastern and Central Europe in 2015: magnitude, temporal trends and projections

Background Within a surveillance of the prevalence and causes of vision impairment in high-income regions and Central/Eastern Europe, we update figures through 2015 and forecast expected values in 2020. Methods Based on a systematic review of medical literature, prevalence of blindness, moderate and severe vision impairment (MSVI), mild vision impairment and presbyopia was estimated for 1990, 2010, 2015, and 2020. Results Age-standardised prevalence of blindness and MSVI for all ages decreased from 1990 to 2015 from 0.26% (0.10–0.46) to 0.15% (0.06–0.26) and from 1.74% (0.76–2.94) to 1.27% (0.55–2.17), respectively. In 2015, the number of individuals affected by blindness, MSVI and mild vision impairment ranged from 70 000, 630 000 and 610 000, respectively, in Australasia to 980 000, 7.46 million and 7.25 million, respectively, in North America and 1.16 million, 9.61 million and 9.47 million, respectively, in Western Europe. In 2015, cataract was the most common cause for blindness, followed by age-related macular degeneration (AMD), glaucoma, uncorrected refractive error, diabetic retinopathy and cornea-related disorders, with declining burden from cataract and AMD over time. Uncorrected refractive error was the leading cause of MSVI. Conclusions While continuing to advance control of cataract and AMD as the leading causes of blindness remains a high priority, overcoming barriers to uptake of refractive error services would address approximately half of the MSVI burden. New data on burden of presbyopia identify this entity as an important public health problem in this population. Additional research on better treatments, better implementation with existing tools and ongoing surveillance of the problem is needed.

Spatial mapping with Gaussian processes and nonstationary Fourier features

The use of covariance kernels is ubiquitous in the field of spatial statistics. Kernels allow data to be mapped into high-dimensional feature spaces and can thus extend simple linear additive methods to nonlinear methods with higher order interactions. However, until recently, there has been a strong reliance on a limited class of stationary kernels such as the Matérn or squared exponential, limiting the expressiveness of these modelling approaches. Recent machine learning research has focused on spectral representations to model arbitrary stationary kernels and introduced more general representations that include classes of nonstationary kernels. In this paper, we exploit the connections between Fourier feature representations, Gaussian processes and neural networks to generalise previous approaches and develop a simple and efficient framework to learn arbitrarily complex nonstationary kernel functions directly from the data, while taking care to avoid overfitting using state-of-the-art methods from deep learning. We highlight the very broad array of kernel classes that could be created within this framework. We apply this to a time series dataset and a remote sensing problem involving land surface temperature in Eastern Africa. We show that without increasing the computational or storage complexity, nonstationary kernels can be used to improve generalisation performance and provide more interpretable results.

Prevalence and causes of vision loss in Latin America and the Caribbean in 2015: magnitude, temporal trends and projections

Objective To estimate the prevalence and causes of blindness and vision impairment for distance and near in Latin America and the Caribbean (LAC) in 2015 and to forecast trends to 2020. Methods A meta-analysis from a global systematic review of 283 cross-sectional, population-representative studies from published and unpublished sources from 1980 to 2014 in the Global Vision Database included 17 published and 6 unpublished studies from LAC. Results In 2015, across LAC, age-standardised prevalence was 0.38% in all ages and 1.56% in those over age 50 for blindness; 2.06% in all ages and 7.86% in those over age 50 for moderate and severe vision impairment (MSVI); 1.89% in all ages and 6.93% in those over age 50 for mild vision impairment and 39.59% in all ages and 45.27% in those over 50 for near vision impairment (NVI). In 2015, 117.86 million persons were vision impaired; of those 2.34 million blind, 12.46 million with MSVI, 11.34 million mildly impaired and 91.72 million had NVI. Cataract is the most common cause of blindness. Undercorrected refractive-error is the most common cause of vision impairment. Conclusions These prevalence estimates indicate that one in five persons across LAC had some degree of vision loss in 2015. We predict that from 2015 to 2020, the absolute numbers of persons with vision loss will increase by 12% to 132.33 million, while the all-age age-standardised prevalence will decrease for blindness by 15% and for other distance vision impairment by 8%. All countries need epidemiologic research to establish accurate national estimates and trends. Universal eye health services must be included in universal health coverage reforms to address disparities, fragmentation and segmentation of healthcare

Prevalence and causes of vision loss in North Africa and Middle East in 2015: magnitude, temporal trends and projections

Background To assess the prevalence and causes of vision impairment in North Africa and the Middle East (NAME) from 1990 to 2015 and to forecast projections for 2020. Methods Based on a systematic review of medical literature, the prevalence of blindness (presenting visual acuity (PVA) <3/60 in the better eye), moderate and severe vision impairment (MSVI; PVA <6/18 but ≥3/60) and mild vision impairment (PVA <6/12 but ≥6/18) was estimated for 2015 and 2020. Results The age-standardised prevalence of blindness and MSVI for all ages and genders decreased from 1990 to 2015, from 1.72 (0.53–3.13) to 0.95% (0.32%–1.71%), and from 6.66 (3.09–10.69) to 4.62% (2.21%–7.33%), respectively, with slightly higher figures for women than men. Cataract was the most common cause of blindness in 1990 and 2015, followed by uncorrected refractive error. Uncorrected refractive error was the leading cause of MSVI in the NAME region in 1990 and 2015, followed by cataract. A reduction in the proportions of blindness and MSVI due to cataract, corneal opacity and trachoma is predicted by 2020. Conversely, an increase in the proportion of blindness attributable to uncorrected refractive error, glaucoma, age-related macular degeneration and diabetic retinopathy is expected. Conclusions In 2015 cataract and uncorrected refractive error were the major causes of vision loss in the NAME region. Proportions of vision impairment from cataract, corneal opacity and trachoma are expected to decrease by 2020, and those from uncorrected refractive error, glaucoma, diabetic retinopathy and age-related macular degeneration are predicted to increase by 2020.

Prevalence and causes of vision loss in South-east Asia and Oceania in 2015: magnitude, temporal trends and projections

Background To assess prevalence and causes of vision impairment in South-east Asia and Oceania regions from 1990 to 2015 and to forecast the figures for 2020. Methods Based on a systematic review of medical literature, prevalence of blindness (presenting visual acuity (PVA) <3/60 in the better eye), moderate and severe vision impairment (MSVI; PVA <6/18 but ≥3/60), mild vision impairment (PVA <6/12 but ≥6/18) and near vision impairment (>N5 or N8 in the presence of normal vision) were estimated for 1990, 2010, 2015 and 2020. Results The age-standardised prevalence of blindness for all ages and both genders was higher in the Oceania region but lower for MSVI when comparing the subregions. The prevalence of near vision impairment in people≥50 years was 41% (uncertainty interval (UI) 18.8 to 65.9). Comparison of the data for 2015 with 2020 predicts a small increase in the numbers of people affected by blindness, MSVI and mild VI in both subregions. The numbers predicted for near VI in South-east Asia are from 90.68 million in 2015 to 102.88 million in 2020. The main causes of blindness and MSVI in both subregions in 2015 were cataract, uncorrected refractive error, glaucoma, corneal disease and age-related macular degeneration. There was no trachoma in Oceania from 1990 and decreasing prevalence in South-east Asia with elimination predicted by 2020. Conclusions In both regions, the main challenges for eye care come from cataract which remains the main cause of blindness with uncorrected refractive error the main cause of MSVI. The trend between 1990 and 2015 is for a lower prevalence of blindness and MSVI in both regions.

Poisson intensity estimation with reproducing kernels

Despite the fundamental nature of the inhomogeneous Poisson process in the theory and application of stochastic processes, and its attractive generalizations (e.g. Cox process), few tractable nonparametric modeling approaches of intensity functions exist, especially when observed points lie in a high-dimensional space. In this paper we develop a new, computationally tractable Reproducing Kernel Hilbert Space (RKHS) formulation for the inhomogeneous Poisson process. We model the square root of the intensity as an RKHS function. Whereas RKHS models used in supervised learning rely on the so-called representer theorem, the form of the inhomogeneous Poisson process likelihood means that the representer theorem does not apply. However, we prove that the representer theorem does hold in an appropriately transformed RKHS, guaranteeing that the optimization of the penalized likelihood can be cast as a tractable finite-dimensional problem. The resulting approach is simple to implement, and readily scales to high dimensions and large-scale datasets.