[1] Drowning: the leading cause of unintentional death in the US
The facts paint a stark picture: each year, about 4,000 people in the United States die from drowning, making it the leading cause of unintentional death from injury among all age groups, and the second-leading cause of unintentional death from injury among children ages 1-14, behind auto accidents. More children ages 1-4 succumb to drowning than any other cause of death except birth defects. About 25% of drowning deaths occur in swimming pools, with children comprising 60% of the victims.
In terms of demographics, males account for nearly 80% of people who die from drowning. Many factors might contribute to higher rates of drowning among males, including increased exposure to water and more risk-taking behavior.
Among nonfatal drowning events, brain damage and other serious consequences are common, including long-term disability. Such incidents result in long-term health problems and costly hospital stays to a much greater extent than fatal incidents. For every child who dies from drowning, another eight receive emergency care for non-fatal drowning incidents. More than 40% of drownings treated in emergency rooms require hospitalization or transfer to specialized units for further care, compared with 8% for all other unintentional injuries.
In the European Union (EU), aquatic risk is similarly severe, with over 6,000 fatalities each year. While in the US aquatic fatalities occur at a rate of 1 per 82,000 residents per year, in the EU it is slightly lower at 1 per 74,000 residents per year. Like the US, drowning is the second-leading cause of unintentional death from injury among children in the EU. For every child who drowns, an additional 140 children are hospitalized for near-drowning. There are, on average, 648 significant swimming pool injuries per day in the EU.
In terms of the financial impact of drowning, several factors are involved. Direct costs, such as medical and legal, are straightforward to analyze. Typical medical costs for a near-drowning victim can range from $75,000 for initial treatment to $180,000 per year for long-term care. The total cost of a single near-drowning that results in brain injury can be more than $4.5 million. The total annual lifetime cost of drownings among children ages 14 and under is approximately $6.8 billion, with children ages 4 and under accounting for $3.4 billion, or half, of these figures. The economic value of each unintentional death from injury is estimated at $790,000 and the comprehensive cost is $2.8 million. The indirect costs of the short- and long-term trauma associated with such incidents, be it in a home/residential setting (depression and loss of livelihood) or a public/commercial one (damaged reputation, negative press, and increased insurance costs), are harder to analyze. And finally, the cost of aquati c injuries to the economy in terms of lost wages and GDP shrinkage is treated using the Value of a Statistical Life (VSL) figure, estimated at $2.4 million per incapacitated person.
[2] The good news: drowning is preventable
As opposed to other risk factors and hazards which may inflict instant damage in both residential and commercial settings, the aquatic incident timeline takes a surprisingly long while to escalate from preliminary risk and early signs of distress to catastrophic permanent damage.
If rescued between 30 to 90 seconds after the water distress incident occurs, swimmers have a near 100% chance of complete recovery with no permanent damage. This time period is the key to providing AI a sufficient window to detect and analyze a developing situation, trigger an alert, and still allow enough time for a human responder to perform a physical intervention. At around two minutes of submersion, a drowning victim has a 94% survival rate if recovered and CPR and artificial respiration are performed properly. After three minutes of submersion, the heart may stop, and at four minutes, irreversible brain damage begins. At around 10 minutes of submersion, a drowning victim has a 14% survival rate if recovered and CPR and artificial respiration are performed properly. Unfortunately, this latter group will usually suffer moderate to severe brain injury.
In over 85% of both fatal and non-fatal drowning cases, an adult is present either on site or at a distance such that they can provide immediate assistance, but does not do so due to lack of information about or awareness of the unfolding incident. This narrow one-minute rescue window, which offers the highest chance of preventing injury, paves the way to better aquatic risk risk management using AI. Since a lifesaving intervention process can take up to 30 seconds, this leaves a 30-second timeframe for AI-driven automatic decision-making systems to assess a situation and trigger an alert for help.
[3] Humans lack the continuous attention essential to managing aquatic risks
Human beings excel at many things: creativity, innovation, and spontaneous thought. Our minds are best suited for thinking forward and switching between ideas built upon our knowledge and past experiences, shuffling between past, present, and future. We tend to get bored, irritated, and drained when performing repetitive and mundane tasks. A 2005 study published in Nature showed that humans perform alarmingly poorly when trying to detect objects or events (e.g., concealed weapons in airport luggage, tumors in mammograms, etc.) which occur at low frequencies. In this study, “low frequency” is a 1% target prevalence level, at which humans miss 25-40% of true targets, a shockingly high level. These are situations where the human is looking at an image containing the sought-after target, but their brains do not register the target and they do not “see” it. In an aquatic environment, distress situations are rarer than this 1% level. Furthermore, many complex and interdependent factors affect vigilance–environmental, physiological, and cognitive–and aquatic environments, which tend to be hot, noisy, and bright, create a challenging setting. A 2001 study looking at commercial pools in the US with professional lifeguards showed the average detection time of a submerged mannequin was one minute and 14 seconds, and in 14% of the 500 trials, the mannequin was not detected within a three-minute rescue window.
The type of attention required for managing aquatic risks requires the ability to perform very mundane, repetitive operations in searching for well-known signs of distress in a challenging environment, all without losing focus. Humans can perform such tasks for short periods of time, but they are easily distracted. Computers and AI are much better suited for this type of task, supporting humans and filling the crucial when they fail to notice an important trigger event.
Furthermore, recent years have seen a steady decline in persons qualified and able to work as lifeguards. This is especially affecting public and municipal facility types, where this shortage of lifeguards is causing facilities to reduce opening hours in some cases and prevents facilities from opening entirely in other cases. The main cause of this phenomena is the low appeal of human lifeguarding due to harsh working conditions combined with very high personal liability.
[4] How AI can dramatically improve aquatic safety
Why now?
In recent years, we have seen an unprecedented boost in performance and proliferation of applications in the fields of Deep Learning-based Artificial Intelligence (DL/AI). These applications are transforming not only high-tech fields, but traditional industries as well. Numerous DL/AI programs, systems, and services are changing the way people work in agriculture, construction, medicine, and sports, to name a few. For many tasks, AI’s level of understanding and decision-making power has reached and surpassed that of humans, and is either assisting or completely replacing humans in specific situations.
The root cause of this Deep Learning revolution can be linked to a combination of algorithmic computer science breakthroughs, hardware and cloud computing advances, and the ever-growing availability of low-cost, high-quality imaging sensors and cameras. Algorithmically, the two seminal Deep Learning papers which mark the modern era of convolutional neural networks are AlexNet (8 layers, published in 2012) and YOLO (329 layers, published in 2016). In terms of hardware, graphical processing units (GPUs) which operate in the tens of GigaFlops range became widespread around 2015, enabling training state-of-the-art networks on large datasets. Together with the widespread availability of scalable, on-demand computer servers, the hardware finally caught up with the newest algorithms. The last key enabler was the price decrease in CMOS sensors and optical hardware which allowed for HD cameras costing roughly $100 each to be easily deployed across residential and commercial facilities.
Why are swimming pools the perfect launchpads for DL/AI?
When water meets optics, several unique imaging problems are created. Phenomena such as reflections, glare, polarization effects, and refraction when transitioning between mediums can hinder visibility in certain situations. Swimming pools, where the water is maintained with low turbidity and is usually less than three meters deep, are the least problematic of all aquatic environments. Furthermore, many pools already have CCTV or security cameras, both in residential and commercial environments, and those which are not can be easily outfitted. This makes pools an ideal launchpad for integrating advanced DL/AI.
What does a DL/AI service looks like?
The following are a few examples of a DL/AI for real-time risk assessment and swimmer analytics operating in a swimming pool environment. The convolutional network used is Lynxight’s proprietary Water Analysis Model ©, which has an RCNN Semantic Segmentation architecture, includes FPN for multiscale performance, and is trained using millions of pool images.
[5] Predictive analytics: rise to the next level of AI
The tactical value of AI in aquatic environments is clear. It is perhaps the critical means by which swimmer distress or drowning events can be detected quickly and with very high probability, halting the escalation of such incidents before they reach the severe phase.
But this is only the beginning. AI can be harnessed for far more advanced strategic applications, analyzing occurrences in real time from a multitude of locations, assessing risk simultaneously for millions of swimmers daily, and gaining insight into the preliminary signs of an imminent safety risk event. As the aquatic dataset grows, complex connections and inconspicuous signals can be detected and linked to future actions and behaviors that put swimmers at risk. Such online dynamic risk assessment, which tracks exposures and is capable of generating a predictive risk model, is the key to truly and meaningfully reducing aquatic risk levels. As predictive and preventative capabilities improve, drowning incidents will be eradicated. This strategic progression from tactical, real-time alerts is analogous to the difference between preventing car accidents by equipping vehicles with good brakes versus employing an advanced early warning collision avoidance system.
All stakeholders in the aquatic risk ecosystem – facility owners, patrons, and insurers – will benefit from this AI-induced risk reduction. As water venues become safer to visit and enjoy, the prevalence of drowning events and related injuries decreases. Financially, as this risk and liability are reduced, the insurance loss ratio goes down, which in turn causes insurance premium levels to go down as well.
