How is Artificial Intelligence Used in Predictive Analytics?

During this time, we have witnessed remarkable progress in analytics automation using artificial intelligence. Deep learning and deep reinforcement learning techniques have also advanced, allowing us to learn, identify, and categorize pictures, text, and voice with high accuracy. All of this has resulted in significant analytics in medical imaging, computer vision applications, text mining, and speech recognition.

The logical next question is how much of this AI development can be applied to sensor data analytics for preventative healthcare systems and predictive maintenance. Can it predict when a machine would break, for instance, using information from a vibration sensor installed on the machine? Or can it accurately screen for early start of cardiac issues using ECG sensor data, allowing steps to be given to avoid further deterioration, and potentially even bringing about improvement? As it turns out, the solution is not that straightforward. Current AI advancements are centered on two major difficulties that limit its wider implementation, particularly in vital sectors such as healthcare and predictive maintenance.

Artificial Intelligence in Predictive Analysis

There is a fundamental distinction between AI systems that need to identify machine failure from vibration data versus systems that recognize people from pictures and those that need to forecast cardiac conditions from ECG data. The former all work with human-generated data, which has an inherent structure, but the latter doesn't. In other words, the earlier systems are driven by a linguistic structure that people can comprehend, but the later ones are not. "There's no way you can build a human-like AI system that doesn't have language at the center of it," says Josh Tenenbaum, professor of cognitive science and computing at MIT. It's one of the most visible characteristics that distinguish human intelligence." A game with visuals, and speech can be said to contain this intrinsic structure known as "language" — simply because they were developed by humans. However, when dealing with raw sensor data created organically by a machine or human body, such a structure may not be easily discernable, making deep learning systems difficult to draw value from it.

Deep learning systems are excellent at detecting abnormalities from sensor data, provided they have learned the models from earlier data. They are, however, not yet adept at developing prognostic/diagnostic systems because of the lack of the underlying 'language' that defines a malfunction or illness.

Interpretability of Artificial Intelligence

Another concern with such predictive diagnostic systems is that they are sensitive and a wrong diagnosis might cost significant expenses. As a result, the models used/learned by AI systems must be physically interpretable by the present scientific knowledge base.

Features/models used to forecast machine breakdown from vibration sensor data, for example, should be explainable and interpretable using physics rules, whereas features/models used to predict cardiac issues from ECG sensor data should employ biology and medical science concepts. Otherwise, even if they deliver outstanding results, certifying and deploying them in multiple scenarios will be problematic.

Supervised machine learning models display extraordinary prediction skills, according to the recent ICML workshop on Human Interpretability in Machine Learning. But can you rely on your model? Will it function during deployment? What more does it have to say about the world? We want models to be interpretable as well as good. Nonetheless, the problem of interpretation appears to be under-specified.

We must realize that without interpretability, we cannot establish ownership and accountability for incorrect diagnosis/prognosis, which is critical in the provided use cases of human and machine health. We also cannot create systems that have been certified by domain experts after repeated trials and tests.