The Katabi lab at MIT’s CSAIL seems to be on a quest to prove that anything a camera can do, a radio can do better. We have recently covered the lab’s work in re-identifying people and tracking sleep position. Their latest work, just published in Nature Medicine, takes a similar approach to tackle the problems associated with self-administration of medications.
Failure to take medications as prescribed contributes heavily to unnecessary hospitalizations and leads to hundreds of thousands of deaths annually. To improve medication adherence it is necessary to not only ensure that patients take their medications, but to also make sure that they administer those medications correctly. Up to 70% of patients do not take insulin as directed, and over 50% use inhalers incorrectly.
To determine if a patient is taking a medication as directed, typically, a clinician will observe them as they self-administer the drug. While this works in the clinic, most administrations will take place at home, without such supervision. To fill this glaring gap, the researchers built a small box that can be mounted on a wall. The box emits low-power radio waves in the WiFi frequency range, which are reflected off of the patient. The reflected signals are analyzed by a machine learning algorithm that has been trained to recognize the motions associated with administration of certain drugs. The system can generate an alert if the patient does not follow the proper protocol.
Experiments were conducted to assess the performance of the system. Excellent results were obtained, with an area under the curve (AUC) of 0.992 for detecting inhaler use, and an AUC of 0.967 for detecting insulin pen use.
Because a radio is used as the sensing mechanism, rather than a camera, the device is much more likely to be accepted as non-intrusive by a patient for in-home use. Moreover, since the device is contactless and always passively monitoring, it places no additional burden on the patient.
The algorithm currently is only able to detect proper use of insulin pens and inhalers, but in principle, with additional training data, could be adjusted to watch for proper use of other medications as well.