Modelling in fm uenza-like illness using online search Vasileios Lampos Computer Science , UCL @lampos lampos.net www
Mapping online search to fm u estimates 12 10 ILI percentage 8 6 4 2 0 2004 2005 2006 2007 2008 Year
Why estimate fm u rates from online search? • Complement traditional syndromic surveillance ‣ timeliness ‣ broader demographic coverage, larger cohort ‣ broader geographic coverage ‣ not a fg ected by closure days ‣ lower cost • Applicable to locations that la cl an established healthcare system
Google Flu Trends — discontinued — popularising an established idea Ginsberg et al . (2009); Eysenbach (2006); Polgreen et al . (2008)
Google Flu Trends — why did it fail ? CDC 0.09 Google Flu Trends 0.08 0.07 0.06 ILI rates 0.05 0.04 0.03 0.02 0.01 0 Jan. '09 Jan. '10 Jan. '11 Jan. '12 Jan. '13 rsv — 25% ILI rate � β 0 � β 1 ⨉ Q , fm u symptoms — 18% where Q is the average benzonatate — 6% symptoms of pneumonia — 6% query frequency upper respiratory infection — 4%
Google Flu Trends — why did it fail ? CDC 0.09 Google Flu Trends 0.08 0.07 0.06 ILI rates 0.05 0.04 0.03 0.02 0.01 0 Jan. '09 Jan. '10 Jan. '11 Jan. '12 Jan. '13 • non-ideal query selection , model simplicity • inappropriate evaluation ( less than 1 fm u season! )
Multivariate, nonlinear, generative models • Treat single search queries as distinct variables • Model nonlinearities 0.08 Raw data Linear fit 0.07 0.06 0.05 ILI rate ILI rates 0.04 0.03 0.02 0.01 0 0 1 2 3 4 5 6 Frequency of query 'how long is flu contagious' -7 frequency of “ how long is fm u contagious ”
Multivariate, nonlinear, generative models • Treat single search queries as distinct variables • Model nonlinearities • Model groups of queries that share common temporal pa tu erns Gaussian Processes (GPs) — distribution over functions that can explain the data — allow some room for model interpretability — can model uncertainty
Correcting the de fj ciencies of Google Flu Trends CDC 0.09 Gaussian Process 0.08 0.07 0.06 ILI rates 0.05 0.04 0.03 0.02 0.01 0 Jan. '09 Jan. '10 Jan. '11 Jan. '12 Jan. '13 • 42% mean absolute error reduction compared to Google Flu Trends • .95 Pearson correlation ( previously .89) with CDC
Modelling uncertainty CDC 0.09 Gaussian Process 0.08 0.07 0.06 ILI rates 0.05 0.04 0.03 0.02 0.01 0 Jan. '09 Jan. '10 Jan. '11 Jan. '12 Jan. '13
Combining GPs with autoregression (AR) CDC 0.09 Gaussian Process with AR 0.08 0.07 0.06 ILI rates 0.05 0.04 0.03 0.02 0.01 0 Jan. '10 Jan. '11 Jan. '12 Jan. '13 • 1 week delay in incorporating historical CDC estimates • 27% mean absolute error reduction over GFT with AR • 52% mean absolute error reduction over GP without AR • .99 Pearson correlation with CDC
Qv ery selection based on meaning • Select search queries based on their semantic similarity to the topic of fm u • Make this possible by using word embeddings , i . e . word representations in a common vector space — learn them using a corpus of 215 million tweets
Qv ery selection based on meaning • Select search queries based on their semantic Analogy: A ( is to ) → B what X ( is to ) → ? similarity to the topic of fm u Rome → Italy London → [ UK , Denmark, Sweden] • Make this possible by using word embeddings , i . e . go → went do → [ did , doing, happened] word representations in a common vector space Messi → football Lebron → [ basketball , bball, NBA] — learn them using a corpus of 215 million tweets Elvis → Presley Aretha → [ Franklin , Ru ffj n, Vandross] UK → Brexit Greece → [ Grexit , Syriza, Tsipras] UK → Farage USA → [ Trump , Farrage, Putin]
Qv ery selection based on meaning • Select search queries based on their semantic similarity to the topic of fm u • Make this possible by using word embeddings , i . e . word representations in a common vector space — learn them using a corpus of 215 million tweets • Combine temporal correlation with semantic similarity ( hybrid similarity ) for optimal feature selection
Qv ery selection based on meaning — Results RCGP (England) 35 Correlation-based feature selection 30 25 ILI rates 20 15 10 5 0 2013 2014 2015 Examples of spurious selected queries prof. surname (70%) name surname recipes (21%) name surname (27%) blood game (12.3%) heating oil (21%) swine fm u vaccine side e fg ects (7.2%)
Qv ery selection based on meaning — Results RCGP (England) 35 Hybrid feature selection 30 25 ILI rates 20 15 10 5 0 2013 2014 2015 • 12.3% performance improvement • .913 Pearson correlation with RCGP ILI rates
i-sense fm u ( Flu Detector ) ! fludetector.cs.ucl.ac.uk @isenseflu
i-sense fm u ( Flu Detector ) • daily fm u estimates for England, publicly accessible • transferred to Public Health England (PHE) • its estimates have been included in the two most recent annual fm u reports of PHE ( gov.uk/ government/statistics/annual-flu-reports ) • open source , github.com/UCL/fludetector-flask • credit to David Guzman for constantly re fj ning it ! fludetector.cs.ucl.ac.uk @isenseflu
Forecasting fm u rates — Ongoing work RCGP (England) 3-weeks ahead forecasts (preliminary model) 50 ILI rates (per 100,000) 40 30 20 10 0 Jan. '15 Jan. '16 Jan. '17 Jan. '18 mean absolute error = 2.56 (cases per 100,000) r = .901 led by Simon Moura
Forecasting fm u rates (US) — Ongoing work 8 CDC (US) 3-weeks ahead forecasts (preliminary model) 7 6 ILI rates (%) 5 4 3 2 1 0 Jan. '15 Jan. '16 Jan. '17 Jan. '18 mean absolute error = 0.33% r = .927 led by Simon Moura
Multi-task learning for fm u Multi-task learning (MTL) vs. single-task learning (STL) • learns models jointly instead of independently • for related tasks is performing be tu er than STL solutions • provides good performance with fewer training samples Flu models with MTL • limit performance loss under sporadic training data • improve accuracy ‣ of regional models within a country ‣ across di fg erent countries
Modelling fm u across US regions with MTL surveil- WA o-fold. Firstly, MT ME ND various ge- VT OR MN NH ID countries — can SD WI NY MA WY MI CT RI to assist IA PA NE NJ NV OH eillance MD UT IL IN DE CO WV CA data. We ex- VA KS MO KY multi-task NC TN AZ OK cess, and NM AR SC formulations. We use MS AL GA eriments on TX LA health and FL AK indicate national mod- Region 1 Region 2 Region 3 Region 4 Region 5 absolute HI Region 6 Region 7 Region 8 Region 9 Region 10 reduced Train 10 US regional models for fm u jointly
MTL across US and US regions Performance for US — 1 year of training data single-task learning multi-task learning 0.88 0.85 0.51 0.44 Pearson correlation mean absolute error
MTL across US and US regions Performance for US regions — 1 year of training data single-task learning multi-task learning 0.87 0.84 0.54 0.47 Pearson correlation mean absolute error
MTL across US and US regions Performance for US regions — 1 year of training data 50% of the data lost single-task learning multi-task learning 0.82 0.77 0.59 0.48 Pearson correlation mean absolute error
MTL across US and England Performance for England — 1 year of training data single-task learning multi-task learning 0.98 0.88 0.85 0.59 Pearson correlation mean absolute error
Why estimate fm u rates from online search? • Complement traditional syndromic surveillance ‣ timeliness ‣ broader demographic coverage, larger cohort ‣ broader geographic coverage ‣ not a fg ected by closure days ‣ lower cost • Applicable to locations that la cl an established healthcare system ‣ oxymoron: healthcare data is required for training the models!
Transfer learning for fm u modelling Main task • train a model for a source location where historical syndromic surveillance data is available • transfer it to a target location where syndromic surveillance data is not available or, in our experiments , ignored Transfer learning steps 1. Learn a regression model for a source location 2. Map search queries from the source to the target domain 3. Transfer the source regression weights to the target domain
Mapping source to target queries • Direct translation does not work • Two similarity components ‣ Semantic similarity ( meaning ) using cross-lingual word embedding representations ( Θ s ) ‣ Temporal similarity based on their frequency time series ( Θ c ) • Joint similarity : Θ = γΘ s + (1 − γ ) Θ c , γ ∈ [0,1]
Source: US, Target: France How similar are their fm u rates? 5 US FR 4 ILI rates (z-scored) 3 2 1 0 -1 2008 2009 2010 2011 2012 2013 2014 2015 2016
Source: US, Target: France MAE = 61.5 , r = .835 MAE = 46.8 , r = .956 MAE = 34.1 , r = .959
Source: US, Target: Australia How similar are their fm u rates? 5 US AU 4 ILI rates (z-scored) 3 2 1 0 -1 2008 2009 2010 2011 2012 2013 2014 2015 2016
Source: US, Target: Australia MAE = 42.6 , r = .7 MAE = 30.3 , r = .915 MAE = 22 , r = .921
Recommend
More recommend