AQAI

Long-term exposure to fine particulate matter (especially PM2.5, i.e. particles with a diameter less than 2.5m) are estimated to cause 8 million excess deaths annually. The impact of air pollution is felt more acutely by the young, with 1 in every 4 deaths under 5 years related to environmental risks. Globally, 93% of children live in places where air pollution levels exceed World Health Organization (WHO) guidelines, defined as PM2.5 values above an annual mean  of 10 ?g/m³ or a 24-hour mean of 25 ?g/m³. Our air pollution problem costs 1% of global GDP - that is $2.6 trillion annually. Unfortunately, due to COVID-19, shifting patterns of travel behavior have meant there is a lack of an accurate way to monitor fluctuating air quality levels globally, leading to uncertainty among public health professionals on children’s exposure to air pollutants. This makes it difficult for humanitarian organizations, and local partners to monitor air pollution accurately, resulting in insufficient evidence to justify investing in respiratory health improvement projects.

There is a need for public health professionals, humanitarian organizations, governments and health insurance providers to monitor air pollution accurately and assess its health impact on populations. At Formative Resilience Knowhow, we believe every child has a right to know what they’re breathing. To do this, we develop AQAI, a machine learning solution to accurately assess the global children populations’ exposure to high levels of air pollution. We generate global predictions of the PM2.5 concentrations (particles with a diameter less than 2.5 micrometers), to augment accurate but limited air pollution data collected using ground sensors with machine learning predictions. Our API, SDK and geovisualization engine enables humanitarian organizations, governments and health insurance providers to monitor air pollution accurately and assess the health impact of air pollution on child populations.

Our solution serves regional governments and NGOs, enabling them to prioritize areas and population for pollution-related investments. After demonstrating our initial proof-of-concept tool to 21 senior scientists, government officials and policy advisors at UNICEF, we established three key impacts AQAI will deliver:

• UNICEF’s regional offices can use AQAI to measure the direct impact of air quality interventions on the reduction in child exposure to high levels of air pollution.
• UNICEF’s health experts indicated that the model would enable an association between COVID-19 incidences and poor air quality and suggested further embedding administrative data from public health systems pre- and post-COVID-19.
• UNICEF’s global field presence could help regional governments standardize the assessment of pollution exposure in order to expand the prioritization of resources to populations most in need

To deliver impact, we intend to work directly with UNICEF Regional Offices in Belize, Costa Rica, and Ecuador. These offices intend to focus on child environmental health under a new UNICEF initiative called “Healthy Environments for Healthy Children”, however, they lack the data to correctly understand pollution exposure. We are working with these offices to deliver important regional datasets on pollution exposure through a dashboard, to engage communities and empower regional governments to ask for funding to improve respiratory services.

Our competitive advantage lies in the unique ability for the team to tackle this challenge.

My previous experience building a unified cloud-based analysis pipeline to support real-time monitoring for London’s Air Quality network has enabled me to  architect the machine learning pipeline required to serve pollution predictions globally via API. During 2020, I moved to Calcutta, India as a pro-bono technical lead and data science consultant. Choked by the fumes from constant traffic and DIY roadside recycling plants incinerating industrial waste, I led a team of five Data Scientists alongside experts at UNICEF to analyze exposure to pollution before and during lockdown events. Experiencing such environmental challenges first-hand imbued a sense of urgency to study environmental inequality with data and statistical techniques. 

As for my cofounder, Prithviraj is the CEO of Formative Resilience Knowhow, and is a patent pending PhD and Fulbright fellow who has studied cost-effective urban air quality measurement techniques extensively. A serial entrepreneur; he previously co-founded a technology start-up dealing with an alternative communication technology for post-disaster management. This was adjudicated by the Indian government as one of the top ten start-ups in the country that works in the field of disaster management. Mr. Pramanik has worked on the densest real-time air quality sensor network in the US, deployed across Chicago.

When I co-founded AQAI back in 2020, my cofounder and I were driven by: 

1. a sense of urgency that threats to children’s respiratory health in developing nations require solutions that can scale, and
2. a belief in the power of data science to predict the respiratory health impact of pollution on children. My cofounder and I used these principles to demonstrate that over 75 million children were exposed to dangerous levels of air pollution during COVID-19 lockdowns.

If we are to harness the enormous potential that technology has in accelerating environmental health analytics, there is no better institution to develop my skills as a technical and entrepreneurial leader than within the context of MIT’s culture of innovation. Alongside a Masters in Urban Science, MIT Solve will help me develop Human Capital, such as developing a board of directors to solicit advice from, and leveraging MIT’s talent network to hire a front end developer to deliver our datasets through accessible, context-dependent visualizations. AQAI’s solution fits with MIT Solve’s Global Challenge: building affordable, accessible, and high-quality health systems that serve everyone, everywhere, as we classify the pollution exposure of 1.8 billion children globally in order to redistribute health services in an equitable and impact-led way. We believe nine months of personalized support from MIT Solve staff’s combined excellence in healthcare and technology will help us expand AQAI to global partners motivated by environmental-health solutions.

Long-term exposure to fine particulate matter is estimated to cause ∼8 million excess deaths annually, and results in economic costs 1% of global GDP - USD 2.6 trillion annually (OECD 2016). Previously, governments and NGOs have used sparse and inaccurate pollution sensors  to inform funding for national health services. This evidence is insufficient to justify investing in air quality improvement projects, leading to an inequitable distribution of health services, which has significantly increased global health inequality. Now due to COVID-19, shifting patterns of travel behaviour means historic pollution data (pre-2019) do not represent post-2020) pollution patterns. leading to uncertainty among public health professionals on children’s exposure to air pollutants. This makes it difficult for humanitarian organisations, governments and health insurance providers to monitor air pollution accurately and assess its health impact on populations.

Imagine if we could monitor air pollution and the exposure of child populations across regions where current data is inaccurate. At AQAI, we develop a machine learning solution to accurately assess the global children populations’ exposure to high levels of air pollution. We predict the concentration of PM2.5 globally, to augment accurate but limited air pollution data collected using ground sensors with machine learning (ML) predictions. 

Let me walk you through a use case. Manuel, a Programme manager at UNICEF’s regional offices in Lima Peru, would like to evaluate the impact of air pollution interventions on children’s health. Although he is aware that there are several Air pollution monitoring stations used in the city, he does not know how to extract the data from them and analyse the results to demonstrate health impacts. Here is where AQAI step in. AQAI aggregates high-quality streams of air pollution data from multiple sources and satellite data from high-resolution imagery, and apply machine learning to predict historical and recent pollution concentrations.

Manuel can combine these historic predictions with socio-economic and demographic datasets, to generate exposure maps. Manuel can now explore which of the local air pollution interventions he deployed (such as pedestrianisation or fines on burning high-carbon fuel sources), significantly lowered pollution levels and reduced exposure. He can export these reports to evidence that health improvements were seen as a result of the techniques used.

Within the next 1 year:

• Product development:
  • To develop an API endpoint for our global level model, to enable a user to request a prediction based on a location point PM2.5 and NO2 concentrations using the pre-trained ensemble model.     
• Piloting and user testing:
  • To develop our first partnership with UNICEF’s regional offices in Belize, Ecuador and Costa Rica (shortlisted locations based on data availability and current scheme of air quality intervention programmes).
  • MIT Solve will enable us to cultivate the network us to form strategic partnerships with partner-funded companies such as qAira & PlumeLabs to gather local ground-based measurements of PM2.5 and NO2 in - areas they are deploying sensors. This will enable us to develop locally tuned models to predict real-time regional air quality.
  • Develop regional, "fine-tuned" versions of the model with local PM2.5 concentration to understand the variability of air pollution patterns across regions, especially with sparse ground sensors. To complete this milestone, we believe UNICEF's global field presence could encourage the participation of regional offices to provide sensor data in regions that have sparse observational data.
• User acquisition/growth
  • To develop our 3-5 partnership with regional offices within our first year, and our first subscription to our API.
  • To map the global model predictions to standard air quality classifications to enable more impact on downstream tasks like exposure measurement and public health analytics. To do this, we aim to incorporate administrative data from public health systems into the API to compare levels of air quality in geographic areas pre- and post-COVID-19 to help distinguish child cases of air pollution-related respiratory conditions from COVID-19 cases.
  • Our current solution predicts weekly averaged PM2.5 concentrations. We want to incorporate Air Quality Standards to indicate the severity of exposure over a fixed period of time. We believe this will enable organizations such as the WHO to update their baseline and historic evidence of pollution levels in developing urban environments (such as Delhi, India where the most recent WHO report uses PM2.5 concentration from 1995).

Within the next 5 years:

• Product development:
  • To be the go-to-provider of environmental data providing a suite of environmental and health-related API endpoints at global coverage.
• Piloting and user testing:
  • Be UNICEF's number one provider of environmental and health data which they use to structure their programmes globally
• User acquisition/growth

  
  

  • Be initiating partnerships with Governmental organisations on directing environment, health and land use policy
  • Starting our first major partnership with a global health insurance provider to drive their environmental health related products

AQAI addressed multiple SDGs:

Good health and Wellbeing

• Mortality rate attributed to ambient air pollution

• Sustainable cities and communities

• We focus on measuring weekly averaged levels of fine particulate matter (e.g. PM2.5 and PM10) at 1 x 1 km globally and 300m X 300m for regional partners.

Affordable and clean energy

• the % change in flaring and combustion activities related to oil, gas and electricity generation and the associated % change in methane PM2.5 and NO2 concentrations.
• The % change in flaring and combustion activities related to oil, gas and electricity generation and the associated % change in methane PM2.5 and NO2 concentrations.

Climate action

• Total greenhouse gas emissions per year (we are focusing on PM10, a short lived climate pollutant that is damaging to respiratory health).

Existing problem

• Globally, 93% of children live in places where air pollution levels exceed World Health Organization (WHO) guidelines, defined as PM2.5 values above 10??/?³ annual mean and 25??/?³ 24-hour mean.

Desired state

• Reduce pollution in areas where child populations are exposed to ambient pollution above the WHO recommended limits and
• Provide sufficient respiratory health investment in areas of high exposure such that there is no excess child mortality related to respiratory health.

People who may suffer from your change.

• Alternative funding for communities that would have received the investment had it not been reprioritised to address respiratory health
• Communities who suffer from “multiple morbidity” which means that investment in respiratory health will not address other health issues such as communicable diseases or non-communicable diseases associated with other sources of infection, e.g. water contamination

Activity timeline1-3 months

Inputs:

• Generate a Global Model of PM2.5 Measurement at 1Kmx1Km scale trained every 2 weeks, with RMSE =<25

Outputs:

• An open source python library hosted on a pypi server that processes open geospatial and satellite data and predicts global PM2.5 concentrations 
• An open source github repository with all code to generate the training data for model training.
• A cloud-hosted API endpoint that, when called, will predict a PM2.5 concentration for a location point.
• A cloud hosted model registry that contains the open source machine learning model for predicting PM2.5 concentrations

Outcomes:

• A technical team member will be able to make a request to an API and receive a PM2.5 prediction for a specific location, and timestamp.
• A technical team member will be able to pull and download the open source repository and download the code that downloads and processes data needed for model training to develop PM2.5 predictions.
• A technical team member will be able to visit a model registry and download a pre-trained model to predict global PM2.5 concentrations.

3-6 months

Inputs:

• Generate a Local Model of PM2.5 concentrations at a local granularity of 300mx300m scale trained every one month (or alternative retraining schedule based on the needs of the regional office) aiming to attain 5% better accuracy over the global model.

Outputs:

• A cloud-hosted API endpoint that when called, will predict the Air Quality Index in a specific location
• Updated cloud hosted global model predicting PM2.5 with improved prediction accuracy as additional training parameters are included
• Cloud-hosted local models predicting PM2.5 with additional local air pollution data from at least UNICEF Regional Office
• Cloud-hosted API endpoints that, when called, will deliver a more accurate pollution prediction for at least one UNICEF Regional Office

Outcomes:

• At least one UNICEF regional office team member will be onboarded to use the global air pollution model and prediction API
• A technical team member will be able to make a request to an API and download the air quality index for a specific location and week
• A technical team member will be able to specify that they would like to receive a request from a local prediction model, and receive a PM2.5 concentration prediction for the region of interest for at least one UNICEF regional office
• A technical team member will be able to download a code repository from github to help download, process and train a local PM2.5 prediction.

6-9 months

Inputs:

• Generate a Global Model of NO2 Measurement at 1Km x 1Km scale trained every 2 weeks, with RMSE =<25
• Onboard 3+ Regional Offices (shortlisted offices are Belize, Costa Rica and Ecuador) interested in monitoring Air Quality & Health such that,
• A member of a UNICEF regional offices will be able to click on a geolocation on a map and select a week and receive an air pollution prediction (PM2.5, NO2 and AQI) for a given location

Outputs:

• A cloud-hosted API endpoint that when called, will predict the NO2 in a specific location
• A cloud-hosted model that is trained to predict global NO2 concentrations in a specific location
• A front end geovisualisation engine hosting the global predictions for PM2.5 and NO2 An open source python library hosted on a pypi server that processes open geospatial and satellite data and predicts global NO2 concentrations

Outcomes:

• A technical team member will be able to make a request to an API and download the NO2 concentration for a specific location and week
• An technical team member will be able to access the trained NO2 model from a model registry
• An technical team member will be able to download a code repository from github to help download, process and train a global NO2 prediction.

9-12 months

Inputs:

• Partnership of 3+ Regional Offices (Shortlisted Regional Offices are Belize, Costa Rica, and Ecuador) such that,
• As a non-technical user I want to be able to see health related data for a location of interest through the geovisualization webapp

Outputs:

• Webhosted geovisualisation tool to show the model predictions alongside additional health datasets
• API endpoint that, when called, exposes the count of populations exposed to high PM2.5 concentrations
• API endpoint that, when called, shows the health impact of the PM2.5 concentrations API endpoint that, when called, produces explainability statistics for the model predictions.

Outcomes:

• As a non-technical user I able to see health related data for a location of interest through the geovisualisation webapp
• As a developer I am able to query an API to see important health related information for a location of interest
• As a developer I am able to understand the most important factors contributing to a prediction of PM2.5 and NO2 via and API

To monitor air pollution across regions where ground sensors are sparse, there is a need to develop a solution to accurately assess the global children populations’ exposure to high levels of air pollution. We developed a machine learning model that predicts the concentration of PM2.5 globally, to augment accurate but limited air pollution data collected using ground sensors with machine learning (ML) predictions. Our solution is delivered using a geographic visualization engine containing layers showing the child population density, and overlays polygons indicating geographic regions with PM2.5 concentrations above the WHO recommended limits (PM2.5 exposure of 10?g/m³ annual mean or 25 ?g/m³ 24-hour mean).

Preprint paper available: https://arxiv.org/abs/2103.12505

Late breaking work (CHI 2021) Video presentation:

Earlier in my career, my former boss and the company CEO professed I was hired as a data scientist because I was a woman. Since then, I have doubted my technical and leadership ability. Despite being a research scientist at the UK’s National Institute for AI and Data Science, my greatest fear about pursuing the entrepreneurial path is that people will not see me as “qualified” enough to lead. Because of this, it has always been a priority of mine to build inclusive, diverse and equitable teams, and this practice is already evident from AQAI’s method of selecting funders and partners. We received valuable guidance from UNICEF an organization working to build programmes to improve accessibility of solutions for children in danger. It is encouraging to see that accessibility and diversity is a shared priority of MIT Solve, as 64 percent of previous teams are women-led. Furthermore, it is exciting that throughout the Solve programme 20 cross-sector leaders will be able to provide alternative perspectives on improving equitable health outcomes as part of our Challenge area.

Beneficiary Segments

• AQAI’s technology produces highly accurate pollution predictions for areas with a good coverage of air pollution sensors to enable UNICEF Multiple Regional Offices monitoring Air Quality exposure and child health impacts
• We aim to introduce our localised ML prediction API and geovisualisation tool to enable better reporting on historical pollution interventions to identify most at-risk locations to be prioritised when introducing healthcare and pollution interventions for local governments to use. This will enhance Air Quality & Health Impact Studies to enable regional governments to prioritize environmental and health initiatives to target pollution-related illness and enable Ministries Land Use to priorities areas for decontamination.
• We aim to provide a “prediction-as-a-service” solution to deliver air pollution metrics, individual exposure levels (over annual average and 24-hour average) as well as health impacts to health insurance providers to improve the accuracy of the models used to calculate health insurance premiums for Health insurance providers

Social and Customer Value Proposition

• The number of workdays lost to air pollution-related illness in 2015 was $2.1 billion, increasing to $3.7 billion in 2060. Demographic groups identified by our solution enables could see a 25% relative increase in productivity, resulting in £500 million in annual savings.
•  Our pollution prediction will identify areas less polluted areas for globally competitive crop yields. A reduction in crop yields as a result of dirty air will weigh on most countries’ economies. Less affected agricultural regions identified through AQAI could result in improved regional export competitiveness and thus economic gains, resulting in $2.7 billion annual savings globally. 
• Incorporating air pollution prediction and associated exposure and health metrics into insurance premiums. The price people are willing to pay each year to not have their health impaired is rising from less than $500 per person in 2015 to as much as $2,800 in 2060. Health insurance providers will want to understand the incidence of pollution exposure to inform their respiratory health models.

Impact Measures

• Collect ground data from schools to improve air quality prediction for humanitarian organisations.
• Incorporate air pollution prediction and associated exposure and health metrics into insurance premiums in Latin America and India.
• Incorporate air pollution prediction and associated exposure and health metrics into civic government for health-first policy planning.

Surplus

• Increase the Number of team members from 1 modelling expert and 1 field deployment expert to 5 Modelling Experts and 20 Field Deployment Experts in our high priority regions
• Further Invest on local sensor networks 
• Posthoc analysis and local collaboration based on the analysis     Invest in interested organization within our regions of interest to build credibility and using the pilots to get customers onboard.

We intent to opt for Service subsidization and Organizational support business models for our Governmental and NGO partners. We will also provide API subscriptions to individual users.

To develop 3-5 partnerships with regional offices and the first subscription to our API.

• We have focussed our efforts so far on developing a proof-of-concept model to predict PM2.5 values globally to attract interest from a range of UNICEF’s regional offices (Belize, Costa Rica and Ecuador). 

To map global model predictions to standard air quality classifications to enable more impact on downstream tasks like exposure measurement and public health analytics. 

• To do this, we aim to incorporate administrative data from public health systems and Air Quality Standards to indicate the severity of exposure over a fixed period of time.

Onboard at least 5 paying clients for our API subscription.

• We are opting for a pilot Deployment with multiple Regional Offices and a free 6 month demo to Field Health Officers, to solicit feedback, attract new customers and build credibility
• We will customize an API request package and front-end dashboard subscription cost access in line with the received feedback.

£15,000 from Subak Climate Fellowship

$100,000 from UNICEF Innovation Fund (final round pending approval)