Incidents happen to all of us. Most of the time, they do not lead to accidents. You can bicker even about the definition of incidents. Take driving a car as an example. You swerve a bit inside the lane you’re driving. This is not an incident, but let’s call it a near miss. You swerve more and you get for a moment in the lane of the drivers coming the opposite way. An incident? Maybe if it happened by accident, but if you avoided a deer coming on your lane and nothing came towards you, it is less of an incident. To make it an accident, something has to follow-out of it as well: an opposing driver you crash into?!

Simple scenarios already demonstrate the challenges in managing and working with risks. The world of risk management is much more complex. With more and more dependencies, complex sourcing strategies and many components in the services and products we design, challenges increase rapidly. And, as can be expected, risk management is not the only party in an organization screaming for resources!


In case of high impact accidents, it is often a pile-up of smaller incidents or near-misses which together coalescence into something dramatic. Often people claim the famous butterfly effect that can cause a perfect storm. However, what is not taken into consideration is that a multitude of other parameters needs to cooperate with the flapping of the butterfly to get the storm.

With accidents it is the same. It is a pile-up of smaller and bigger things. Although it seems an accident happens out of the blue, with hindsight it always turns out it was an accident waiting to happen. The signs were there if you know where to look. Knowing where to look is exactly the issues. Ideally, you want to look forward and have suggestion where the likelihood of an accident is the highest. For this, you want to get as many parameters as possible and not be limited on reporting by people.


Major accidents are a cascade of different events coming together to push a situation over the edge. To increase prevention of major accidents, machine learning can be a big support while limiting the need for increasing many resources in risk management. With machine learning a large and diverse set of parameters can be analyzed to predict possible accidents. Also, clear feedback loops are available to make the predictions better over time as new incidents and accidents

  • New Incidents: Feedback for predictions what is might become an incidents or an incident prone situation.
  • Accidents: Feedback for predictions what might case a real accident

The advantage in risk management is that not all incidents lead to accidents so already based on incidents prediction models can improve over time with machine learning. This is especially important is the number of real accidents are inherently rare. (At least they should be). The advantage of machine learning are:

  • Start moving away from the binary situation: Accident/No Accident or Incident/No Incident and more towards probability which allow better maximize use of available resources and people.
  • In addition, a more holistic view can be taken on preventing accidents where interdependencies can be better modeled to predict hot-spots for potential accidents and incidents
  • The implicit knowledge of risk managers and auditors can be capture through human/machine collaboration. Especially in risk management this is important as over the years much implicit experience is build-up by risk managers and they are a small professional group resulting to relative high risks when they leave.


An example to leverage machine learning is in health care. There is a complex network of people involved from specialists to GP to nurses both in terms of Care and Cure. Incidents are often viewed personal failures and patients regularly feel they have to fight the system to get the cure they deserv.

Take the example of a young 10-year-old child. Her GP, over time, missed four time the diagnose appendicitis. Naturally, as the child got more ill, parents get more frustrated. At one point it became so acute she had to go to hospital. It turned into an acute appendicitis where the appendix burst and the child fought for her life.

By taking the diagnoses less binary at monitor/alert that a situation is getting more critical earlier support can be designed in the process such as second-opinion above a certain risk level. This way it is not the parents having to push on the GP which creates a conflict situation. Also the GP is not threatened in his professional capabilities as it is based on automatic prediction. And last, the situation de-escalate to what might be life threatening due to earlier intervention. It helps prevent build-up of human tension by making it more objective based on technology.


In the end there is no need to wait for accidents to happen. Leveraging human/machine collaboration can go a long way even when situations and inter-dependencies get more complex. We expect through new technologies based on Machine Learning that the role of auditors, risk managers and risk management will step-up to a know level.

Blog by Mike van Pamelen