Imperial College London

DrSethFlaxman

Faculty of Natural SciencesDepartment of Mathematics

Senior Lecturer in Statistical Machine Learning
 
 
 
//

Contact

 

s.flaxman

 
 
//

Location

 

522Huxley BuildingSouth Kensington Campus

//

Summary

 

Publications

Publication Type
Year
to

60 results found

Vollmer M, Radhakrishnan S, Kont M, Flaxman S, Bhatt S, Costelloe C, Honeyford C, Aylin P, Cooke G, Redhead J, White P, Ferguson N, Hauck K, Nayagam AS, Perez Guzman PNet al., 2020, Report 29: The impact of the COVID-19 epidemic on all-cause attendances to emergency departments in two large London hospitals: an observational study

The health care system in England has been highly affected by the surge in demand due to patients afflicted by COVID-19. Yet the impact of the pandemic on the care seeking behaviour of patients and thus on Emergency department (ED) services is unknown, especially for non-COVID-19 related emergencies. In this report, we aimed to assess how the reorganisation of hospital care and admission policies to respond to the COVID-19 epidemic affected ED attendances and emergency hospital admissions. We performed time-series analyses of present year vs historic (2015-2019) trends of ED attendances between March 12 and May 31 at two large central London hospitals part of Imperial College Healthcare NHS Trust (ICHNT) and compared these to regional and national trends. Historic attendances data to ICHNT and publicly available NHS situation reports were used to calibrate time series auto-regressive integrated moving average (ARIMA) forecasting models. We thus predicted the (conterfactual) expected number of ED attendances between March 12 (when the first public health measure leading to lock-down started in England) to May 31, 2020 (when the analysis was censored) at ICHNT, at all acute London Trusts and nationally. The forecasted trends were compared to observed data for the same periods of time. Lastly, we analysed the trends at ICHNT disaggregating by mode of arrival, distance from postcode of patient residence to hospital and primary diagnosis amongst those that were subsequently admitted to hospital and compared these data to an average for the same period of time in the years 2015 to 2019.During the study period (January 1 to May 31, 2020) there was an overall decrease in ED attendances of 35% at ICHNT, of 50% across all London NHS Trusts and 53% nationally. For ICHNT, the decrease in attendances was mainly amongst those aged younger than 65 and those arriving by their own means (e.g. personal or public transport). Increasing distance (km) from postcode of residence to hospi

Report

Lavezzo E, Franchin E, Ciavarella C, Cuomo-Dannenburg G, Barzon L, Del Vecchio C, Rossi L, Manganelli R, Loregian A, Navarin N, Abate D, Sciro M, Merigliano S, De Canale E, Vanuzzo MC, Besutti V, Saluzzo F, Onelia F, Pacenti M, Parisi S, Carretta G, Donato D, Flor L, Cocchio S, Masi G, Sperduti A, Cattarino L, Salvador R, Nicoletti M, Caldart F, Castelli G, Nieddu E, Labella B, Fava L, Drigo M, Gaythorpe KAM, Imperial College COVID-19 Response Team, Brazzale AR, Toppo S, Trevisan M, Baldo V, Donnelly CA, Ferguson NM, Dorigatti I, Crisanti Aet al., 2020, Suppression of a SARS-CoV-2 outbreak in the Italian municipality of Vo'., Nature, ISSN: 0028-0836

On the 21st of February 2020 a resident of the municipality of Vo', a small town near Padua, died of pneumonia due to SARS-CoV-2 infection1. This was the first COVID-19 death detected in Italy since the emergence of SARS-CoV-2 in the Chinese city of Wuhan, Hubei province2. In response, the regional authorities imposed the lockdown of the whole municipality for 14 days3. We collected information on the demography, clinical presentation, hospitalization, contact network and presence of SARS-CoV-2 infection in nasopharyngeal swabs for 85.9% and 71.5% of the population of Vo' at two consecutive time points. On the first survey, which was conducted around the time the town lockdown started, we found a prevalence of infection of 2.6% (95% confidence interval (CI) 2.1-3.3%). On the second survey, which was conducted at the end of the lockdown, we found a prevalence of 1.2% (95% Confidence Interval (CI) 0.8-1.8%). Notably, 42.5% (95% CI 31.5-54.6%) of the confirmed SARS-CoV-2 infections detected across the two surveys were asymptomatic (i.e. did not have symptoms at the time of swab testing and did not develop symptoms afterwards). The mean serial interval was 7.2 days (95% CI 5.9-9.6). We found no statistically significant difference in the viral load of symptomatic versus asymptomatic infections (p-values 0.62 and 0.74 for E and RdRp genes, respectively, Exact Wilcoxon-Mann-Whitney test). This study sheds new light on the frequency of asymptomatic SARS-CoV-2 infection, their infectivity (as measured by the viral load) and provides new insights into its transmission dynamics and the efficacy of the implemented control measures.

Journal article

Okell LC, Verity R, Watson OJ, Mishra S, Walker P, Whittaker C, Katzourakis A, Donnelly CA, Riley S, Ghani AC, Gandy A, Flaxman S, Ferguson NM, Bhatt Set al., 2020, Have deaths from COVID-19 in Europe plateaued due to herd immunity?, Lancet, Vol: 395, Pages: e110-e111

Journal article

Flaxman S, Mishra S, Gandy A, Unwin HJT, Mellan TA, Coupland H, Whittaker C, Zhu H, Berah T, Eaton JW, Monod M, Perez Guzman PN, Schmit N, Cilloni L, Ainslie K, Baguelin M, Boonyasiri A, Boyd O, Cattarino L, Cucunuba Perez Z, Cuomo-Dannenburg G, Dighe A, Djaafara A, Dorigatti I, van Elsland S, Fitzjohn R, Gaythorpe K, Geidelberg L, Grassly N, Green W, Hallett T, Hamlet A, Hinsley W, Jeffrey B, Knock E, Laydon D, Nedjati Gilani G, Nouvellet P, Parag K, Siveroni I, Thompson H, Verity R, Volz E, Walters C, Wang H, Watson O, Winskill P, Xi X, Walker P, Ghani AC, Donnelly CA, Riley SM, Vollmer MAC, Ferguson NM, Okell LC, Bhatt Set al., 2020, Estimating the effects of non-pharmaceutical interventions on COVID-19 in Europe, Nature, ISSN: 0028-0836

Following the emergence of a novel coronavirus1 (SARS-CoV-2) and its spread outside of China, Europe has experienced large epidemics. In response, many European countries have implemented unprecedented non-pharmaceutical interventions such as closure of schools and national lockdowns. We study the impact of major interventions across 11 European countries for the period from the start of COVID-19 until the 4th of May 2020 when lockdowns started to be lifted. Our model calculates backwards from observed deaths to estimate transmission that occurred several weeks prior, allowing for the time lag between infection and death. We use partial pooling of information between countries with both individual and shared effects on the reproduction number. Pooling allows more information to be used, helps overcome data idiosyncrasies, and enables more timely estimates. Our model relies on fixed estimates of some epidemiological parameters such as the infection fatality rate, does not include importation or subnational variation and assumes that changes in the reproduction number are an immediate response to interventions rather than gradual changes in behavior. Amidst the ongoing pandemic, we rely on death data that is incomplete, with systematic biases in reporting, and subject to future consolidation. We estimate that, for all the countries we consider, current interventions have been sufficient to drive the reproduction number Rt below 1 (probability Rt< 1.0 is 99.9%) and achieve epidemic control. We estimate that, across all 11 countries, between 12 and 15 million individuals have been infected with SARS-CoV-2 up to 4th May, representing between 3.2% and 4.0% of the population. Our results show that major non-pharmaceutical interventions and lockdown in particular have had a large effect on reducing transmission. Continued intervention should be considered to keep transmission of SARS-CoV-2 under control.

Journal article

Dighe A, Cattarino L, Cuomo-Dannenburg G, Skarp J, Imai N, Bhatia S, Gaythorpe K, Ainslie K, Baguelin M, Bhatt S, Boonyasiri A, Boyd O, Brazeau N, Charles G, Cooper L, Coupland H, Cucunuba Perez Z, Djaafara A, Dorigatti I, Eales O, Eaton J, van Elsland S, Ferreira Do Nascimento F, Fitzjohn R, Flaxman S, Fraser K, Geidelberg L, Green W, Hallett T, Hamlet A, Hauck K, Haw D, Hinsley W, Jeffrey B, Knock E, Laydon D, Lees J, Mellan T, Mishra S, Nedjati Gilani G, Nouvellet P, Okell L, Parag K, Pons Salort M, Ragonnet-Cronin M, Thompson H, Unwin H, Verity R, Whittaker C, Whittles L, Xi X, Ghani A, Donnelly C, Ferguson N, Riley Set al., 2020, Report 25: Response to COVID-19 in South Korea and implications for lifting stringent interventions, 25

While South Korea experienced a sharp growth in COVID-19 cases early in the global pandemic, it has since rapidly reduced rates of infection and now maintains low numbers of daily new cases. Despite using less stringent “lockdown” measures than other affected countries, strong social distancing measures have been advised in high incidence areas and a 38% national decrease in movement occurred voluntarily between February 24th - March 1st. Suspected and confirmed cases were isolated quickly even during the rapid expansion of the epidemic and identification of the Shincheonji cluster. South Korea swiftly scaled up testing capacity and was able to maintain case-based interventions throughout. However, individual case-based contact tracing, not associated with a specific cluster, was a relatively minor aspect of their control program, with cluster investigations accounting for a far higher proportion of cases: the underlying epidemic was driven by a series of linked clusters, with 48% of all cases in the Shincheonji cluster and 20% in other clusters. Case-based contacts currently account for only 11% of total cases. The high volume of testing and low number of deaths suggests that South Korea experienced a small epidemic of infections relative to other countries. Therefore, caution is needed in attempting to duplicate the South Korean response in settings with larger more generalized epidemics. Finding, testing and isolating cases that are linked to clusters may be more difficult in such settings.

Report

Unwin H, Mishra S, Bradley VC, Gandy A, Vollmer M, Mellan T, Coupland H, Ainslie K, Whittaker C, Ish-Horowicz J, Filippi S, Xi X, Monod M, Ratmann O, Hutchinson M, Valka F, Zhu H, Hawryluk I, Milton P, Baguelin M, Boonyasiri A, Brazeau N, Cattarino L, Charles G, Cooper L, Cucunuba Perez Z, Cuomo-Dannenburg G, Djaafara A, Dorigatti I, Eales O, Eaton J, van Elsland S, Fitzjohn R, Gaythorpe K, Green W, Hallett T, Hinsley W, Imai N, Jeffrey B, Knock E, Laydon D, Lees J, Nedjati Gilani G, Nouvellet P, Okell L, Ower A, Parag K, Siveroni I, Thompson H, Verity R, Walker P, Walters C, Wang Y, Watson O, Whittles L, Ghani A, Ferguson N, Riley S, Donnelly C, Bhatt S, Flaxman Set al., 2020, Report 23: State-level tracking of COVID-19 in the United States

our estimates show that the percentage of individuals that have been infected is 4.1% [3.7%-4.5%], with widevariation between states. For all states, even for the worst affected states, we estimate that less than a quarter of thepopulation has been infected; in New York, for example, we estimate that 16.6% [12.8%-21.6%] of individuals have beeninfected to date. Our attack rates for New York are in line with those from recent serological studies [1] broadly supportingour choice of infection fatality rate.There is variation in the initial reproduction number, which is likely due to a range of factors; we find a strong associationbetween the initial reproduction number with both population density (measured at the state level) and the chronologicaldate when 10 cumulative deaths occurred (a crude estimate of the date of locally sustained transmission).Our estimates suggest that the epidemic is not under control in much of the US: as of 17 May 2020 the reproductionnumber is above the critical threshold (1.0) in 24 [95% CI: 20-30] states. Higher reproduction numbers are geographicallyclustered in the South and Midwest, where epidemics are still developing, while we estimate lower reproduction numbersin states that have already suffered high COVID-19 mortality (such as the Northeast). These estimates suggest that cautionmust be taken in loosening current restrictions if effective additional measures are not put in place.We predict that increased mobility following relaxation of social distancing will lead to resurgence of transmission, keepingall else constant. We predict that deaths over the next two-month period could exceed current cumulative deathsby greater than two-fold, if the relationship between mobility and transmission remains unchanged. Our results suggestthat factors modulating transmission such as rapid testing, contact tracing and behavioural precautions are crucial to offsetthe rise of transmission associated with loosening of social distancing. Overall, we

Report

Winskill P, Whittaker C, Walker P, Watson O, Laydon D, Imai N, Cuomo-Dannenburg G, Ainslie K, Baguelin M, Bhatt S, Boonyasiri A, Cattarino L, Ciavarella C, Cooper L, Coupland H, Cucunuba Perez Z, van Elsland S, Fitzjohn R, Flaxman S, Gaythorpe K, Green W, Hallett T, Hamlet A, Hinsley W, Knock E, Lees J, Mellan T, Mishra S, Nedjati Gilani G, Nouvellet P, Okell L, Parag K, Thompson H, Unwin H, Wang Y, Whittles L, Xi X, Ferguson N, Donnelly C, Ghani Aet al., 2020, Report 22: Equity in response to the COVID-19 pandemic: an assessment of the direct and indirect impacts on disadvantaged and vulnerable populations in low- and lower middle-income countries, 22

The impact of the COVID-19 pandemic in low-income settings is likely to be more severe due to limited healthcare capacity. Within these settings, however, there exists unfair or avoidable differences in health among different groups in society – health inequities – that mean that some groups are particularly at risk from the negative direct and indirect consequences of COVID-19. The structural determinants of these are often reflected in differences by income strata, with the poorest populations having limited access to preventative measures such as handwashing. Their more fragile income status will also mean that they are likely to be employed in occupations that are not amenable to social-distancing measures, thereby further reducing their ability to protect themselves from infection. Furthermore, these populations may also lack access to timely healthcare on becoming ill. We explore these relationships by using large-scale household surveys to quantify the differences in handwashing access, occupation and hospital access with respect to wealth status in low-income settings. We use a COVID-19 transmission model to demonstrate the impact of these differences. Our results demonstrate clear trends that the probability of death from COVID-19 increases with increasing poverty. On average, we estimate a 32.0% (2.5th-97.5th centile 8.0%-72.5%) increase in the probability of death in the poorest quintile compared to the wealthiest quintile from these three factors alone. We further explore how risk mediators and the indirect impacts of COVID-19 may also hit these same disadvantaged and vulnerable the hardest. We find that larger, inter-generational households that may hamper efforts to protect the elderly if social distancing are associated with lower-income countries and, within LMICs, lower wealth status. Poorer populations are also more susceptible to food security issues - with these populations having the highest levels under-nourishment whilst also being

Report

Mellan T, Hoeltgebaum H, Mishra S, Whittaker C, Schnekenberg R, Gandy A, Unwin H, Vollmer M, Coupland H, Hawryluk I, Rodrigues Faria N, Vesga J, Zhu H, Hutchinson M, Ratmann O, Monod M, Ainslie K, Baguelin M, Bhatia S, Boonyasiri A, Brazeau N, Charles G, Cooper L, Cucunuba Perez Z, Cuomo-Dannenburg G, Dighe A, Djaafara A, Eaton J, van Elsland S, Fitzjohn R, Fraser K, Gaythorpe K, Green W, Hayes S, Imai N, Jeffrey B, Knock E, Laydon D, Lees J, Mangal T, Mousa A, Nedjati Gilani G, Nouvellet P, Olivera Mesa D, Parag K, Pickles M, Thompson H, Verity R, Walters C, Wang H, Wang Y, Watson O, Whittles L, Xi X, Okell L, Dorigatti I, Walker P, Ghani A, Riley S, Ferguson N, Donnelly C, Flaxman S, Bhatt Set al., 2020, Report 21: Estimating COVID-19 cases and reproduction number in Brazil

Brazil is an epicentre for COVID-19 in Latin America. In this report we describe the Brazilian epidemicusing three epidemiological measures: the number of infections, the number of deaths and the reproduction number. Our modelling framework requires sufficient death data to estimate trends, and wetherefore limit our analysis to 16 states that have experienced a total of more than fifty deaths. Thedistribution of deaths among states is highly heterogeneous, with 5 states—São Paulo, Rio de Janeiro,Ceará, Pernambuco and Amazonas—accounting for 81% of deaths reported to date. In these states, weestimate that the percentage of people that have been infected with SARS-CoV-2 ranges from 3.3% (95%CI: 2.8%-3.7%) in São Paulo to 10.6% (95% CI: 8.8%-12.1%) in Amazonas. The reproduction number (ameasure of transmission intensity) at the start of the epidemic meant that an infected individual wouldinfect three or four others on average. Following non-pharmaceutical interventions such as school closures and decreases in population mobility, we show that the reproduction number has dropped substantially in each state. However, for all 16 states we study, we estimate with high confidence that thereproduction number remains above 1. A reproduction number above 1 means that the epidemic isnot yet controlled and will continue to grow. These trends are in stark contrast to other major COVID19 epidemics in Europe and Asia where enforced lockdowns have successfully driven the reproductionnumber below 1. While the Brazilian epidemic is still relatively nascent on a national scale, our resultssuggest that further action is needed to limit spread and prevent health system overload.

Report

Vollmer M, Mishra S, Unwin H, Gandy A, Melan T, Bradley V, Zhu H, Coupland H, Hawryluk I, Hutchinson M, Ratmann O, Monod M, Walker P, Whittaker C, Cattarino L, Ciavarella C, Cilloni L, Ainslie K, Baguelin M, Bhatia S, Boonyasiri A, Brazeau N, Charles G, Cooper L, Cucunuba Perez Z, Cuomo-Dannenburg G, Dighe A, Djaafara A, Eaton J, van Elsland S, Fitzjohn R, Gaythorpe K, Green W, Hayes S, Imai N, Jeffrey B, Knock E, Laydon D, Lees J, Mangal T, Mousa A, Nedjati Gilani G, Nouvellet P, Olivera Mesa D, Parag K, Pickles M, Thompson H, Verity R, Walters C, Wang H, Wang Y, Watson O, Whittles L, Xi X, Ghani A, Riley S, Okell L, Donnelly C, Ferguson N, Dorigatti I, Flaxman S, Bhatt Set al., 2020, Report 20: A sub-national analysis of the rate of transmission of Covid-19 in Italy

Italy was the first European country to experience sustained local transmission of COVID-19. As of 1st May 2020, the Italian health authorities reported 28; 238 deaths nationally. To control the epidemic, the Italian government implemented a suite of non-pharmaceutical interventions (NPIs), including school and university closures, social distancing and full lockdown involving banning of public gatherings and non essential movement. In this report, we model the effect of NPIs on transmission using data on average mobility. We estimate that the average reproduction number (a measure of transmission intensity) is currently below one for all Italian regions, and significantly so for the majority of the regions. Despite the large number of deaths, the proportion of population that has been infected by SARS-CoV-2 (the attack rate) is far from the herd immunity threshold in all Italian regions, with the highest attack rate observed in Lombardy (13.18% [10.66%-16.70%]). Italy is set to relax the currently implemented NPIs from 4th May 2020. Given the control achieved by NPIs, we consider three scenarios for the next 8 weeks: a scenario in which mobility remains the same as during the lockdown, a scenario in which mobility returns to pre-lockdown levels by 20%, and a scenario in which mobility returns to pre-lockdown levels by 40%. The scenarios explored assume that mobility is scaled evenly across all dimensions, that behaviour stays the same as before NPIs were implemented, that no pharmaceutical interventions are introduced, and it does not include transmission reduction from contact tracing, testing and the isolation of confirmed or suspected cases. We find that, in the absence of additional interventions, even a 20% return to pre-lockdown mobility could lead to a resurgence in the number of deaths far greater than experienced in the current wave in several regions. Future increases in the number of deaths will lag behind the increase in transmission intensity and so a

Report

Cheng C-Y, Wang N, Wong TY, Congdon N, He M, Wang YX, Braithwaite T, Casson RJ, Cicinelli MV, Das A, Flaxman SR, Jonas JB, Keeffe JE, Kempen JH, Leasher J, Limburg H, Naidoo K, Pesudovs K, Resnikoff S, Silvester AJ, Tahhan N, Taylor HR, Bourne RRAet al., 2020, Prevalence and causes of vision loss in East Asia in 2015: magnitude, temporal trends and projections, BRITISH JOURNAL OF OPHTHALMOLOGY, Vol: 104, Pages: 616-622, ISSN: 0007-1161

Journal article

Lisi E, Malekzadeh M, Haddadi H, Lau FD-H, Flaxman Set al., 2020, Modelling and forecasting art movements with CGANs, Publisher: ROYAL SOC

Working paper

Lisi E, Malekzadeh M, Haddadi H, Lau D-H, Flaxman Set al., 2020, Modeling and forecasting art movements with CGANs, Royal Society Open Science, Vol: 7, ISSN: 2054-5703

Conditional generative adversarial networks (CGANs) are a recent and popular method for generating samples from a probability distribution conditioned on latent information. The latent information often comes in the form of a discrete label from a small set. We propose a novel method for training CGANs which allows us to condition on a sequence of continuous latent distributions f(1), …, f(K). This training allows CGANs to generate samples from a sequence of distributions. We apply our method to paintings from a sequence of artistic movements, where each movement is considered to be its own distribution. Exploiting the temporal aspect of the data, a vector autoregressive (VAR) model is fitted to the means of the latent distributions that we learn, and used for one-step-ahead forecasting, to predict the latent distribution of a future art movement f(K+1). Realizations from this distribution can be used by the CGAN to generate ‘future’ paintings. In experiments, this novel methodology generates accurate predictions of the evolution of art. The training set consists of a large dataset of past paintings. While there is no agreement on exactly what current art period we find ourselves in, we test on plausible candidate sets of present art, and show that the mean distance to our predictions is small.

Journal article

Flaxman S, Mishra S, Gandy A, Unwin H, Coupland H, Mellan T, Zhu H, Berah T, Eaton J, Perez Guzman P, Schmit N, Cilloni L, Ainslie K, Baguelin M, Blake I, Boonyasiri A, Boyd O, Cattarino L, Ciavarella C, Cooper L, Cucunuba Perez Z, Cuomo-Dannenburg G, Dighe A, Djaafara A, Dorigatti I, van Elsland S, Fitzjohn R, Fu H, Gaythorpe K, Geidelberg L, Grassly N, Green W, Hallett T, Hamlet A, Hinsley W, Jeffrey B, Jorgensen D, Knock E, Laydon D, Nedjati Gilani G, Nouvellet P, Parag K, Siveroni I, Thompson H, Verity R, Volz E, Walters C, Wang H, Wang Y, Watson O, Winskill P, Xi X, Whittaker C, Walker P, Ghani A, Donnelly C, Riley S, Okell L, Vollmer M, Ferguson N, Bhatt Set al., 2020, Report 13: Estimating the number of infections and the impact of non-pharmaceutical interventions on COVID-19 in 11 European countries

Following the emergence of a novel coronavirus (SARS-CoV-2) and its spread outside of China, Europe is now experiencing large epidemics. In response, many European countries have implemented unprecedented non-pharmaceutical interventions including case isolation, the closure of schools and universities, banning of mass gatherings and/or public events, and most recently, widescale social distancing including local and national lockdowns. In this report, we use a semi-mechanistic Bayesian hierarchical model to attempt to infer the impact of these interventions across 11 European countries. Our methods assume that changes in the reproductive number – a measure of transmission - are an immediate response to these interventions being implemented rather than broader gradual changes in behaviour. Our model estimates these changes by calculating backwards from the deaths observed over time to estimate transmission that occurred several weeks prior, allowing for the time lag between infection and death. One of the key assumptions of the model is that each intervention has the same effect on the reproduction number across countries and over time. This allows us to leverage a greater amount of data across Europe to estimate these effects. It also means that our results are driven strongly by the data from countries with more advanced epidemics, and earlier interventions, such as Italy and Spain. We find that the slowing growth in daily reported deaths in Italy is consistent with a significant impact of interventions implemented several weeks earlier. In Italy, we estimate that the effective reproduction number, Rt, dropped to close to 1 around the time of lockdown (11th March), although with a high level of uncertainty. Overall, we estimate that countries have managed to reduce their reproduction number. Our estimates have wide credible intervals and contain 1 for countries that have implemented all interventions considered in our analysis. This means that the reproducti

Report

Naidoo K, Kempen JH, Gichuhi S, Braithwaite T, Casson RJ, Cicinelli MV, Das A, Flaxman SR, Jonas JB, Keeffe JE, Leasher J, Limburg H, Pesudovs K, Resnikoff S, Silvester AJ, Tahhan N, Taylor HR, Wong TY, Bourne RRA, Vision Loss Expert Group of the Global Burden of Disease Studyet al., 2020, Prevalence and causes of vision loss in sub-Saharan Africa in 2015: magnitude, temporal trends and projections., Br J Ophthalmol

BACKGROUND: This study aimed to assess the prevalence and causes of vision loss in sub-Saharan Africa (SSA) in 2015, compared with prior years, and to estimate expected values for 2020. METHODS: A systematic review and meta-analysis assessed the prevalence of blindness (presenting distance visual acuity <3/60 in the better eye), moderate and severe vision impairment (MSVI; presenting distance visual acuity <6/18 but ≥3/60) and mild vision impairment (MVI; presenting distance visual acuity <6/12 and ≥6/18), and also near vision impairment (<N6 or N8 in the presence of ≥6/12 best-corrected distance visual acuity) in SSA for 1990, 2010, 2015 and 2020.In SSA, age-standardised prevalence of blindness, MSVI and MVI in 2015 were 1.03% (80% uncertainty interval (UI) 0.39-1.81), 3.64% (80% UI 1.71-5.94) and 2.94% (80% UI 1.05-5.34), respectively, for male and 1.08% (80% UI 0.40-1.93), 3.84% (80% UI 1.72-6.37) and 3.06% (80% UI 1.07-5.61) for females, constituting a significant decrease since 2010 for both genders. There were an estimated 4.28 million blind individuals and 17.36 million individuals with MSVI; 101.08 million individuals were estimated to have near vision loss due to presbyopia. Cataract was the most common cause of blindness (40.1%), whereas undercorrected refractive error (URE) (48.5%) was the most common cause of MSVI. Sub-Saharan West Africa had the highest proportion of blindness compared with the other SSA subregions. CONCLUSIONS: Cataract and URE, two of the major causes of blindness and vision impairment, are reversible with treatment and thus promising targets to alleviate vision impairment in SSA.

Journal article

Wolock TM, Flaxman SR, Eaton JW, 2019, Inferring HIV incidence trends and transmission dynamics with a spatio-temporal HIV epidemic model, Publisher: arXiv

Reliable estimation of spatio-temporal trends in population-level HIVincidence is becoming an increasingly critical component of HIV preventionpolicy-making. However, direct measurement is nearly impossible. Current,widely used models infer incidence from survey and surveillance seroprevalencedata, but they require unrealistic assumptions about spatial independenceacross spatial units. In this study, we present an epidemic model of HIV thatexplicitly simulates the spatial dynamics of HIV over many small, interactingareal units. By integrating all available population-level data, we are able toinfer not only spatio-temporally varying incidence, but also ART initiationrates and patient counts. Our study illustrates the feasibility of applyingcompartmental models to larger inferential problems than those to which theyare typically applied, as well as the value of data fusion approaches toinfectious disease modeling.

Working paper

Davis SPX, Kumar S, Alexandrov Y, Bhargava A, da Silva Xavier G, Rutter GA, Frankel P, Sahai E, Flaxman S, French PMW, McGinty Jet al., 2019, Convolutional neural networks for reconstruction of undersampled optical projection tomography data applied to in vivo imaging of zebrafish., Journal of Biophotonics, Vol: 12, ISSN: 1864-063X

Optical projection tomography (OPT) is a 3D mesoscopic imaging modality that can utilize absorption or fluorescence contrast. 3D images can be rapidly reconstructed from tomographic data sets sampled with sufficient numbers of projection angles using the Radon transform, as is typically implemented with optically cleared samples of the mm-to-cm scale. For in vivo imaging, considerations of phototoxicity and the need to maintain animals under anesthesia typically preclude the acquisition of OPT data at a sufficient number of angles to avoid artifacts in the reconstructed images. For sparse samples, this can be addressed with iterative algorithms to reconstruct 3D images from undersampled OPT data, but the data processing times present a significant challenge for studies imaging multiple animals. We show here that convolutional neural networks (CNN) can be used in place of iterative algorithms to remove artifacts - reducing processing time for an undersampled in vivo zebrafish dataset from 77 to 15 minutes. We also show that using CNN produces reconstructions of equivalent quality to CS with 40% fewer projections. We further show that diverse training data classes, for example ex vivo mouse tissue data, can be used for CNN-based reconstructions of OPT data of other species including live zebrafish.

Journal article

Runge J, Nowack P, Kretschmer M, Flaxman S, Sejdinovic Det al., 2019, Detecting and quantifying causal associations in large nonlinear time series datasets, Science Advances, Vol: 5, Pages: 1-15, ISSN: 2375-2548

Identifying causal relationships and quantifying their strength fromobservational time series data are key problems in disciplines deal-ing with complex dynamical systems such as the Earth system orthe human body. Data-driven causal inference in such systems ischallenging since datasets are often high-dimensional and nonlinear with limited sample sizes. Here we introduce a novel method thatflexibly combines linear or nonlinear conditional independence testswith a causal discovery algorithm to estimate causal networks fromlarge-scale time series datasets. We validate the method on timeseries of well-understood physical mechanisms in the climate sys-tem and the human heart and using large-scale synthetic datasetsmimicking the typical properties of real world data. The experi-ments demonstrate that our method outperforms state-of-the-arttechniques in detection power, which opens up entirely new possi-bilities to discover and quantify causal networks from time seriesacross a range of research fields.

Journal article

Flaxman S, Chirico M, Pereira P, Loeffler Cet al., Scalable high-resolution forecasting of sparse spatiotemporal events with kernel methods: a winning solution to the NIJ "Real-Time Crime Forecasting Challenge", Annals of Applied Statistics, ISSN: 1932-6157

We propose a generic spatiotemporal event forecasting method,which we developed for the National Institute of Justice’s (NIJ) RealTime Crime Forecasting Challenge (National Institute of Justice,2017). Our method is a spatiotemporal forecasting model combining scalable randomized Reproducing Kernel Hilbert Space (RKHS)methods for approximating Gaussian processes with autoregressivesmoothing kernels in a regularized supervised learning framework.While the smoothing kernels capture the two main approaches in current use in the field of crime forecasting, kernel density estimation(KDE) and self-exciting point process (SEPP) models, the RKHScomponent of the model can be understood as an approximation tothe popular log-Gaussian Cox Process model. For inference, we discretize the spatiotemporal point pattern and learn a log-intensityfunction using the Poisson likelihood and highly efficient gradientbased optimization methods. Model hyperparameters including quality of RKHS approximation, spatial and temporal kernel lengthscales,number of autoregressive lags, bandwidths for smoothing kernels, aswell as cell shape, size, and rotation, were learned using crossvalidation. Resulting predictions significantly exceeded baseline KDEestimates and SEPP models for sparse events.

Journal article

Nangia V, Jonas J, George R, Lingam V, Ellwein L, Cicinelli MV, Das A, Flaxman SR, Keeffe J, Leasher J, Naidoo K, Pesudovs K, Resnikoff S, Silvester A, Tahhan N, Taylor H, Wong T, Bourne Ret al., 2019, Prevalence and causes of blindness and vision impairment magnitude, Temporal Trends, and Projections in South and Central Asia, British Journal of Ophthalmology, Vol: 103, Pages: 871-877, ISSN: 0007-1161

Background: To assess prevalence and causes of vision loss in Central 68 and South Asia.Methods: A systematic review of medical literature assessed the prevalence of blindness (presenting visual acuity<3/60 in the better eye), moderate and severe vision impairment (MSVI; presenting visual acuity <6/18 but ≥3/60) and mild vision impairment (MVI; presenting visual acuity <6/12 and ≥6/18) in Central and South Asia for 1990, 2010, 2015, and 2020.Results: In Central and South Asia combined, age-standardized prevalences of blindness, MSVI and MVI in 2015 were for men and women 2.80% (80%uncertainty interval (UI):1.14-4.91) and 3.47% (80%UI:1.45-5.99),16.75% (80%UI:5.60-30.84) and 20.06% (80%UI:7.15-36.12), 11.49% (80%UI:3.43-21.44) and 12.77% (80%UI:4.04-23.48), respectively, with a significant decrease in the study period for both gender. In South Asia in 2015, 11.76 million individuals (32.65% of the global blindness figure) were blind and 61.19 million individuals (28.3% of the global total) had MSVI. From 1990 to 2015, cataract (accounting for 36.58% of all cases with blindness in 2015) was the most common cause of blindness, followed by undercorrected refractive error (36.43%), glaucoma (5.81%), age-related macular degeneration (2.44%), corneal diseases (2.43%), diabetic retinopathy (0.16%) and trachoma (0.04%). For MSVI in South Asia 2015, most common causes were under corrected refractive error (accounting for 66.39% of all cases with MSVI), followed by cataract (23.62%), age related macular degeneration (1.31%) and glaucoma (1.09%).Conclusions: One third of the global blind resided in South Asia in 2015, although the age-standardized prevalence of blindness and MSVI decreased significantly between 1990 and 2015.

Journal article

Keeffe J, Casson R, Pesudovs K, Taylor H, Cicinelli M, Das A, Flaxman SR, Jonas J, Kempen J, Leasher J, Limburg H, Naidoo K, Silvester A, Stevens G, Tahhan N, Wong T, Resnikoff S, Bourne Ret al., 2019, Prevalence and causes of vision loss in South-east Asia and Oceania in 2015: magnitude, temporal trends, and projections, British Journal of Ophthalmology, Vol: 103, Pages: 878-884, ISSN: 0007-1161

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-standardized prevalence of blindness for all ages and both genders was higher in the Oceania region but lower for MSVI when comparing the sub-regions. The prevalence of near vision impairment in people ≥50 years was 41% (UI 18.8-65.9). 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 sub-regions. 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 sub-regions 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:Inboth 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 revalence of blindness and MSVI in both regions.

Journal article

Crawford L, Flaxman SR, Runcie DE, West Met al., 2019, Variable Prioritization in Nonlinear Black Box Methods: A Genetic Association Case Study, Annals of Applied Statistics, Vol: 13, Pages: 958-989, ISSN: 1932-6157

The central aim in this paper is to address variable selection questions innonlinear and nonparametric regression. Motivated by statistical genetics,where nonlinear interactions are of particular interest, we introduce a noveland interpretable way to summarize the relative importance of predictorvariables. Methodologically, we develop the "RelATive cEntrality" (RATE)measure to prioritize candidate genetic variants that are not just marginallyimportant, but whose associations also stem from significant covaryingrelationships with other variants in the data. We illustrate RATE throughBayesian Gaussian process regression, but the methodological innovations applyto other "black box" methods. It is known that nonlinear models often exhibitgreater predictive accuracy than linear models, particularly for phenotypesgenerated by complex genetic architectures. With detailed simulations and tworeal data association mapping studies, we show that applying RATE enables anexplanation for this improved performance.

Journal article

Tusting LS, Bisanzio D, Alabaster G, Cameron E, Cibulskis R, Davies M, Flaxman S, Gibson HS, Knudsen J, Mbogo C, Okumu FO, von Seidlein L, Weiss DJ, Lindsay SW, Gething PW, Bhatt Set al., 2019, Mapping changes in housing in sub-Saharan Africa from 2000 to 2015, Nature, Vol: 568, Pages: 391-394, ISSN: 0028-0836

Access to adequate housing is a fundamental human right, essential to human security, nutrition and health, and a core objective of the United Nations Sustainable Development Goals1,2. Globally, the housing need is most acute in Africa, where the population will more than double by 2050. However, existing data on housing quality across Africa are limited primarily to urban areas and are mostly recorded at the national level. Here we quantify changes in housing in sub-Saharan Africa from 2000 to 2015 by combining national survey data within a geostatistical framework. We show a marked transformation of housing in urban and rural sub-Saharan Africa between 2000 and 2015, with the prevalence of improved housing (with improved water and sanitation, sufficient living area and durable construction) doubling from 11% (95% confidence interval, 10-12%) to 23% (21-25%). However, 53 (50-57) million urban Africans (47% (44-50%) of the urban population analysed) were living in unimproved housing in 2015. We provide high-resolution, standardized estimates of housing conditions across sub-Saharan Africa. Our maps provide a baseline for measuring change and a mechanism to guide interventions during the era of the Sustainable Development Goals.

Journal article

Ton J-F, Flaxman S, Sejdinovic D, Bhatt Set al., 2018, Spatial mapping with Gaussian processes and nonstationary Fourier features, Spatial Statistics, Vol: 28, Pages: 59-78, ISSN: 2211-6753

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.

Journal article

Law HC, Sejdinovic D, Cameron E, Lucas T, Flaxman S, Battle K, Fukumizu Ket al., 2018, Variational Learning on Aggregate Outputs with Gaussian Processes, NIPS

Journal article

Hu A, Flaxman S, 2018, Multimodal sentiment analysis to explore the structure of emotions, KDD 2018, Publisher: ACM, Pages: 350-358

We propose a novel approach to multimodal sentiment analysis using deep neural networks combining visual analysis and natural language processing. Our goal is different than the standard sentiment analysis goal of predicting whether a sentence expresses positive or negative sentiment; instead, we aim to infer the latent emotional state of the user. Thus, we focus on predicting the emotion word tags attached by users to their Tumblr posts, treating these as "self-reported emotions." We demonstrate that our multimodal model combining both text and image features outperforms separate models based solely on either images or text. Our model's results are interpretable, automatically yielding sensible word lists associated with emotions. We explore the structure of emotions implied by our model and compare it to what has been posited in the psychology literature, and validate our model on a set of images that have been used in psychology studies. Finally, our work also provides a useful tool for the growing academic study of images - both photographs and memes - on social networks.

Conference paper

Bourne RRA, Jonas JB, Bron AM, Cicinelli MV, Das A, Flaxman SR, Friedman DS, Keeffe JE, Kempen JH, Leasher J, Limburg H, Naidoo K, Pesudovs K, Peto T, Saadine J, Silvester AJ, Tahhan N, Taylor HR, Varma R, Wong TY, Resnikoff Set al., 2018, Prevalence and causes of vision loss in high-income countries and in Eastern and Central Europe in 2015: magnitude, temporal trends and projections, BRITISH JOURNAL OF OPHTHALMOLOGY, Vol: 102, Pages: 575-585, ISSN: 0007-1161

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.

Journal article

Law HCL, Sutherland D, Sejdinovic D, Flaxman SRet al., 2018, Bayesian approaches to distribution regression, The 21st International Conference on Artificial Intelligence and Statistics: AISTATS 2018, Publisher: AISTATS

Distribution regression has recently attractedmuch interest as a generic solution to the problem of supervised learning where labels are available at the group level, rather than at the individual level. Current approaches, however, do notpropagate the uncertainty in observations due tosampling variability in the groups. This effectively assumes that small and large groups areestimated equally well, and should have equalweight in the final regression. We account forthis uncertainty with a Bayesian distribution regression formalism, improving the robustnessand performance of the model when group sizesvary. We frame our models in a neural networkstyle, allowing for simple MAP inference usingbackpropagation to learn the parameters, as wellas MCMC-based inference which can fully propagate uncertainty. We demonstrate our approachon illustrative toy datasets, as well as on a challenging problem of predicting age from images.

Conference paper

Abbati G, Tosi A, Osborne M, Flaxman SRet al., 2018, AdaGeo: adaptive geometric learning for optimization and sampling, 21 st International Conference on Artificial Intelligence and Statistics (AISTAT)S, Publisher: PMLR, Pages: 226-234

Gradient-based optimization and Markov Chain Monte Carlo sampling can be found at the heart of several machine learning methods. In high-dimensional settings, well-known issues such as slow-mixing, non-convexity and correlations can hinder the algorithms’ efficiency. In order to overcome these difficulties, we propose AdaGeo, a preconditioning framework for adaptively learning the geometry of the parameter space during optimization or sampling. In particular, we use the Gaussian process latent variable model (GP-LVM) to represent a lower-dimensional embedding of the parameters, identifying the underlying Riemannian manifold on which the optimization or sampling is taking place. Samples or optimization steps are consequently proposed based on the geometry of the manifold. We apply our framework to stochastic gradient descent, stochastic gradient Langevin dynamics, and stochastic gradient Riemannian Langevin dynamics, and show performance improvements for both optimization and sampling.

Conference paper

Flaxman SR, Bourne RA, Resnikoff S, Ackland P, Braithwaite Tet al., 2017, Global causes of blindness and distance vision impairment 1990–2020: a systematic review and meta-analysis, The Lancet Global Health, Vol: 5, Pages: e1221-e1234, ISSN: 2214-109X

BackgroundContemporary data for causes of vision impairment and blindness form an important basis of recommendations in public health policies. Refreshment of the Global Vision Database with recently published data sources permitted modelling of cause of vision loss data from 1990 to 2015, further disaggregation by cause, and forecasts to 2020.MethodsIn this systematic review and meta-analysis, we analysed published and unpublished population-based data for the causes of vision impairment and blindness from 1980 to 2014. We identified population-based studies published before July 8, 2014, by searching online databases with no language restrictions (MEDLINE from Jan 1, 1946, and Embase from Jan 1, 1974, and the WHO Library Database). We fitted a series of regression models to estimate the proportion of moderate or severe vision impairment (defined as presenting visual acuity of <6/18 but ≥3/60 in the better eye) and blindness (presenting visual acuity of <3/60 in the better eye) by cause, age, region, and year.FindingsWe identified 288 studies of 3 983 541 participants contributing data from 98 countries. Among the global population with moderate or severe vision impairment in 2015 (216·6 million [80% uncertainty interval 98·5 million to 359·1 million]), the leading causes were uncorrected refractive error (116·3 million [49·4 million to 202·1 million]), cataract (52·6 million [18·2 million to 109·6 million]), age-related macular degeneration (8·4 million [0·9 million to 29·5 million]), glaucoma (4·0 million [0·6 million to 13·3 million]), and diabetic retinopathy (2·6 million [0·2 million to 9·9 million]). Among the global population who were blind in 2015 (36·0 million [12·9 million to 65·4 million]), the leading causes were cataract (12·6 million [3·4 million to 28·7 million]), uncorrected

Journal article

Goodman B, Flaxman S, 2017, European Union regulations on algorithmic decision-making and a “Right to Explanation”, AI Magazine, Vol: 38, ISSN: 0738-4602

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.

Journal article

This data is extracted from the Web of Science and reproduced under a licence from Thomson Reuters. You may not copy or re-distribute this data in whole or in part without the written consent of the Science business of Thomson Reuters.

Request URL: http://wlsprd.imperial.ac.uk:80/respub/WEB-INF/jsp/search-html.jsp Request URI: /respub/WEB-INF/jsp/search-html.jsp Query String: respub-action=search.html&id=00770956&limit=30&person=true