Body Weight Readings using Machine Learning on IoT Data

Body Weight Readings using Machine Learning on IoT Data

More than 50 technology changemakers worked with proprietary, anonymized, HIPAA compliant* Smart Home sensor data. The team developed a Machine Learning model that can infer true body weight from an IoT connected scale and assist with the remote sensing of elderly and at-risk patients, improving their overall care with minimal intervention.


The partner EmPowerYu provides resident-authorized formal and informal caregivers with non-intrusive means to be assured of the wellbeing of care recipients and to be alerted in exceptional conditions. Existing signals include quality and time of meals, quality and time of sleep, and general activity of the house including departure and return.


The Problem

Remote monitoring with EmPowerYu helps people at high medical risk live more safely in the community. Most remote monitoring requires the resident to actively engage with smartphones, but older adults and disabled people often have physical, cognitive, technical, economic, or language barriers that make smartphone systems impractical. EmPowerYu works for everyone. EmPowerYu’s Smart Home system is effortless for the resident, so we collect complete and long-term whole-person datasets, unlike wearable sensors that people stop using after a while.

Effective remote monitoring reduces healthcare utilization, improves care decisions, and lowers stress for caregivers. True Body Weight is the actual weight of the body tissues. It is the ‘first thing in the morning’ weight. E.g. (after rising from bed, before dressing in day clothing, after first toileting, before eating or drinking).

True Body Weight changes over an interval of days (for example, one week) can indicate significant changes in wellbeing, such as a worsening of heart disease or decline in appetite, that would benefit from intervention.

A five-pound change of real body weight is clinically significant. Detection is confounded because weight over the course of a day can vary by five pounds because of daily activities that are not clinically meaningful, like the weight of clothing. Reliable adjustment of a weight measurement to real body weight under standard conditions would permit a more accurate and earlier determination of concern.

Excess alarms to care recipients cause alert fatigue and must be avoided to assure timely response and product success.


The Solution

Within two months the team created a solution to clean and normalize biomarker and sensor data provided in various unstructured formats. From that the team engineered 13 predictive features of true body weight, trained and evaluated 20 different machine learning models, deployed the model as a Flask app with an interactive dashboard, created a Flask API with the model productionized and communicable via REST endpoints, and developed a fully coded pipeline for a REST API capable of normalizing the disaggregated sensor data in real-time from inception to inference.


Partner Feedback



Detecting Forest Wood Fire using IoT Sensor Data

Detecting Forest Wood Fire using IoT Sensor Data

50 AI engineers collaborated for 8 weeks to analyze sensor data and test possible systemic data models to develop an intelligent recognition algorithm to detect fire of different types of wood. 

The project partner, Dryad Networks, is a Germany-based startup that provides ultra-early detection of wildfires as well as health and growth-monitoring of forests using solar-powered gas sensors in a large-scale IoT sensor network. Dryad aims to tackle wildfires, which are causing up to 20% of global CO2 emissions and have a devastating impact on biodiversity.


Dryad - Wildfire Detection Germany

The problem

The world’s forests are burning! If current trends continue, up to 170 million hectares could be lost until 2030 and with it, we gradually lose the earth’s great carbon sink consuming 110 billion metric tons of CO2.


The project outcomes

The project’s goal was to build an intelligent model that will detect fire of different types of wood through analysis of the existing sensor data, thereby enabling alarms for firefighters early enough, so they can extinguish it. During the period of eight (8) weeks, the team designed and implemented several data-based pipelines, leveraging the dataset provided by the Dryad team. Combing a massive and extensive analysis of the datasets provided, together with the state-of-the-art machine learning techniques, the team delivered the following.

The results of this project lie in the state of art machine learning models and correctly classify the sensor data into two categories, “in-smoke” and “clean-air”.  The model developed in this project is scalable and replicable. Such a solution has the potential to reduce forest fires, thereby enabling alarms for firefighters early enough, so they can extinguish them. This will ultimately help to achieve the sustainable development goals in the areas of life on land and climate action.