AI / ML to reduce or eliminate DUI Casualties. A Solution.

The Problem:

37 people die in a car accident every day in US due to DUI.

In 2023 there were ~13,500 fatalities due to DUI in the States and this number was a ~25% increase since 2016.

Although technology and processing power exist today to reduce such fatalities that continue to disrupt lives and families, but it is either not utilized or implemented yet.

Imagine a scenario where the car won't start if it identifies the driver under influence.

More and more cars have both internal and external cameras that are able to understand human patterns such as:

1. How alert the driver is.

2. What is the driver's reaction to external environment and how quick.

3. How distracted the driver is with the tech / phone and other functions in the car.

4. The problem can further be alleviated if technology synergies are implemented by auto makers and device makers (cell phones, wrist worn fitness devices, etc). This will permit the cars onboard computer to make an almost perfectly accurate prediction of driver's state of well being (Under Influence OR Normal). Markers of breathing rate, heart rate, blood pressure changes etc. can be taken into consideration along with video footage of the driver.

Such an implementation will reduce DUI accidents and causalities.

One of the Solutions (I understand others or even AI may come up with more elegant solutions in the future if not already have):

Training Data should not pose a challenge to train the system in this scenario. Training Data can be acquired:

1.       Most Police Officers capture videos today when they stop cars and question drivers regarding alcohol / drugs.

2.       Medical / Health industry and hospitals have detailed health and bio marker analysis when they find patients under influence. I.e. Heart Rate, Breathing Rate and Variability, Blood Oxygen (SpO2) and how it affects their reaction times.

3.       Insurance companies leverage Actuarial insights and statistics which can further improve prediction accuracy of this system.   

I would propose the following approach to get most accurate prediction while ensuring speed to market and model explainability:

Assumption:

1.       Large amount of data from disparate sources – as mentioned above.

2.       Data quality will be a challenge.

3.       Data would mostly be structured given the sources of Data mentioned above.  

4.       Data would need to be pre-processed and standardized to ensure accurate prediction and reducing noise and reducing the “over-fitting” problem.

Machine Learning will suffice in this scenario as the need of Feature Engineering is almost moot as the available data sources are pre-labeled by domain experts (Doctors, Police Officers, Statisticians) , and the need for explainability negates the need for Deep Learning and leveraging Neural Network.  

Further improvement in predictions can be made by leveraging GAN (Generative Adversarial Networks) to fine-tune the model.

Although the proposed solution does not leverage NLP (Natural Language Processing) – but if leveraged can be used for speech pattern recognition such as slurring and other minute changes in speech when an individual is under influence.

 Such a technology implementation is relatively easy and cheap to implement and can save thousands of lives and make our roads safer.