This article is a summary of Dr. Oliver’s talk on Data Science and Computational Models for public health, and why we should interconnect them with data stakeholders (policymakers, healthcare workers, etc) in a pandemic. Link to the talk: PathCheck DICE - Global Health Innovators Seminar - 16 - YouTube
Dr. Oliver’s work is primarily based on enabling policymakers to undertake better decisions through data-driven analysis. In particular, her team worked closely with the President of the Valencia region, Spain to help them make better data-driven decisions across multiple waves of the pandemic.
Process of data collection to governmental roles:
Official website of Valencian Data Science Group: https://infocoronavirus.gva.es/es/grup-de-ciencies-de-dades-del-covid-19-de-la-comunitat-valencianaIn May 2020, Spain was enforced with a degradingly strict lockdown, during which the team studied the efficacy of various interventions using survey data as well as aggregated mobility data. Few of the impacts studied are the following :
- The success of the Stay at Home Campaign
On average, 88% (working days) and 92% (weekends) of the population in the Valencia Region have not left their area of residence from 16th March to 27th April 2020
- Analysis of medical services from healthcare departments
Over 80% of all the 24 populations per health department region have remained in said regions
- Labor mobility
On Average, 60% fewer people outside their area of residence during working hours than a working day in November (16 March - 17th April period)
Website of visualization system and variables for the impact of lockdown: https://www.arcgis.com/apps/opsdashboard/index.html#/778023f21baf447cb387c98c318f1e5c
- Algorithms that detect forms of matrices in communities are made to identify mobility communities (which are self-contained regions that are active internally, but not with adjacent regions), named macro health zones.
To continue with interventions for COVID-19 Impacts, epidemiological models (which are various scenarios to predict different outcomes in the future) were investigated to enhance ongoing measures of policies. These policies were enforced to hinder the impact of covid, and therefore it is essential to measure the rate of reduction in covid infections it has caused.
There are 3 common variations used of the epidemiological models:
- SIER Metapopulation model: divides the population into 4 states (Susceptible, Exposed, Infected, Recovered), the transition between each of the states depends on the probabilities which depend on the virus characteristics.
- Agent-based model: Does not model a meta-based population, it is a conglomerate of models. In this case, all 5 million individuals in Valencia are modeled as agents, thus enabling understanding of policies at a granular level.
- Deep learning model : Valencia formulated a COVID-19 infections predictor using the deep learning model, consisting of 2 LSTMs (a recurrent neural network that uses patterns to predict the next likely scenarios):
The model was used since the 3rd COVID-19 wave in the Valencian Region (Valencia is currently experiencing its 6th wave), accurately predicting the 3rd and 6th wave and effects on healthcare strain
Dr. Oliver also talked about the work from her group that focuses on understanding the mental health, economical and behavioral impacts of the pandemic and various interventions on the public. For this, the team made use of crowdsourcing technology such as the Covid-19 Impact Survey: https://covid19impactsurvey.org/
The survey was launched out of the realization of the siloed information regarding:
- Social Contact behavior
- Economic and Labor impact
- Prevalence of symptoms
- Testing availability
- Contact Tracing
General trends found:
- The emotional impacts of lockdown and abusive use of technology viewed in adults was most prevalent in 18-29 females
- Women have also been the most compliant in adopting individual protection measures, they tend to report that they are unable to self-isolate due to the caretaking responsibilities on them, fear of stigmatization and being psychologically impossible
- Indoor ventilation has been the least deployed measure
For the general population, the perception of government methods aligns with the intensity of the COVID-19 waves.
Preliminary results of the survey from 720,000 individuals: https://www.jmir.org/2020/9/e21319/ (Assessing the Impact of the COVID-19 Pandemic in Spain: Large-Scale, Online, Self-Reported Population Survey). Visualized presentation of feedback: https://ellisalicante.org/en/covid19impactsurvey
Key correlations between social isolation, economic impact, protection of measures, and perception of risk were identified with graphical trends:
Conclusion: The COVID-19 Pandemic is not a healthcare issue, but a societal issue. Hence, it requires holistic and multidisciplinary approaches
A holistic approach to tackle the COVID-19 pandemic beyond the use of mobile apps: https://www.vodafone-institut.de/digitising-europe/tracing-the-virus-a-holistic-view/
A virtual cycle needs to be developed between the 3 elements:
Q&A Highlights:
- Between the agent-based model and the deep-learning model, which is the most flexible?
The deep-learning based model. It continues learning from the current data and updates itself. Though the other models are more interpretable, they are also more traditionalist and requiring of more manual tuning of parameters(which adapt slowly from variables such as virus changes, and vaccinations, whilst the deep learning model is self-changing and does not require a consistent update
- How did you manage the understanding of interpretability, particularly from the deep learning model? Since various scientists from Health to Data to Policymakers will define it differently
It is beyond the models, the interpretability, in general, is the significance of the results. The key stakeholders should be the real members of the team, rather than indifferent/passive listeners. While using deep learning based solutions, what is more, important is the mutual understanding of the importance of results and their actionable intents.
- Which non-pharma interventions were the most effective in Valencia?
Insight to 12 interventions: https://public.tableau.com/app/profile/kristina.p8284/viz/Prescriptions_16117279637400/Visualize A machine learning model was used to predict the reproductive rates that enabled us to run the interventions by contributions. It was found that labor and education interventions were the most impactful
- How do you feel about the use of AI for decision-making?
It is true that there may be a point where policy-makers become dependent on AI 'opinion'. But if data is an objective representation of reality, it works indicative to the human decision. AI is meant to help human decisions, not overrule them. AI is for humanity augmentation more than replacement. Al also has limitations, the data used to train the models could be biased from a biased reality, hence they can exacerbate and amplify them. Sometimes the limitations are intrinsic to the models, the lack of transparency because the models are too complex, the models are hackable, brittle, not fool-proof (adversarial machine learning). We are already coexisting with high non-transparent algorithms that are already influencing our decisions. Any form of automation for human machines is called AI.
Relevant resources:
Published paper on leveraging data and technology in the context of pandemics: https://www.vodafone-institut.de/wp-content/uploads/2020/10/VFI-DPA_Fighting_COVID_with_Data_report_2020.pdf
Open Data Science to fight COVID-19: Winning the 500k XPRIZE Pandemic Response Challenge
https://rua.ua.es/dspace/bitstream/10045/119336/2/Lozano_etal_2021_ECML-PKDD_preprint.pdf
Give more data, awareness, and control to individual citizens, and they will help COVID-19 containment
https://link.springer.com/article/10.1007/s10676-020-09572-w
More publications:
https://www.jmir.org/2020/9/e21319/
https://www.vodafone-institut.de/digitising-europe/tracing-the-virus-a-holistic-view/
https://link.springer.com/article/10.1007/s10676-020-09572-w
https://rua.ua.es/dspace/bitstream/10045/119336/2/Lozano_etal_2021_ECML-PKDD_preprint.pdf
https://infocoronavirus.gva.es/es/grup-de-ciencies-de-dades-del-covid-19-de-la-comunitat-valenciana
Thank both our speaker — Dr. Nuria Oliver — and the organizers — Rohan Sukumaran, Graham Dodge, Ramesh Raskar, Nina Reščič, Shanice Hudson, and Tavpritesh Sethi!
Please check out dice.pathcheck.org/talks.html to watch this and previous talks and feel free to join the volunteer-driven Slack workspace of the PathCheck Foundation here: https://tiny.cc/pathcheckslack!