The a-Heal Machine Learning Bandage

Every year, the human population sustains roughly 24 billion wounds, with an average range of one to three per individual. When left untreated, large wounds can result in permanent tissue damage, loss of mobility, and even psychological disturbances. For those without easy access to public healthcare, wound treatment remains a pressing issue. To combat this issue, engineers have created a device only possible in the present era: a bandage that uses artificial intelligence to heal wounds.

Designed by engineers at the University of California–Santa Cruz and the University of California–Davis, a-Heal is a bioelectrical wound therapy platform that utilizes a machine-learning (ML) physician to diagnose and prescribe wound treatments. Led by UC Santa Cruz’s Professor of Electrical and Computer Engineering (ECE) Marco Rolandi, the team developed a wearable device able to photograph the wound, receive instructions from an ML physician and prescribe treatment accordingly.

As a wearable device, featuring a bandage bottom, a-Heal is first placed over the wound site. Once applied, the device captures images and wirelessly transmits them to the ML physician, which then analyzes the photos and determines the wound stage. If the observed state of the wound displays a deterrence from the optimal healing path, the ML physician prescribes treatment as needed. The wearable device then applies this treatment, either fluid or electrotherapy, and its camera module continues to update the ML physician with bihourly images.

The wireless component of a-Heal is made up of several components: a camera to capture images of the wound, dual printed circuit boards (PCBs) for illumination, a microcontroller to translate commands from the ML Physician, a transparent polydimethylsiloxane (PDMS) body for storing treatments, two bioelectronic actuators, and a USB-C port, all protected by a 3D printed waterproof enclosure.

The PCBs surrounding the camera are circular rings consisting of twelve LED banks for lighting. Aside from providing illumination, they also relay the ML Physician’s instructions to the microcontroller, having received them via Wi-Fi. The microcontroller then translates these instructions into low-level signals to be sent to the bioelectronic actuators.

The bioelectronic actuators, which apply treatment, hold another PCB. An iontophoretic pump body houses 8 PDMS reservoirs for said treatment, four of which contain sterile saline for electrotherapy use, the other four containing fluoxetine hydrochloride solution to give fluoxetine a positive charge. There exists also a central reservoir for the counter electrode. All reservoirs have Ag/AgCl electrodes and are attached to the wound by anionic polymer-filled capillary.

Imaging of the wound is a multi-step process, with each monitoring cycle lasting roughly 24 hours. First, the imaging module, which consists of the camera, PCBs, plano-convex lens, and microcontroller, takes 11 images to get an accurate multi-dimensional view of the wound site. These images are then sent wirelessly to the ML Physician by the module, which then sleeps for the two hours allotted before resuming photographing, a process meant to allow the battery to last the full day-long cycle by reducing the amount of power used.

The wearable device’s treatment is equally complex. Housed in the PDMS body, the two available treatments are electric field (EF) therapy and fluoxetine (Flx) injection. EF aids wound healing by moving keratinocytes, cells of the epidermis, towards the wound edge, where they can begin repairing the damaged skin. Flx is a selective serotonin reuptake inhibitor which decreases inflammation as it transitions the wound to rebuilding tissue to fill the wound site.

Upon collecting information, the wearable device transfers it wirelessly to the ML Physician to analyze and prescribe treatment. Headed by Santa Cruz’s Assistant Professor of Applied Mathematics Marcella Gomez, the creation of the ML Physician used reinforcement learning, where the model is tasked with an end goal, and rewarded for its progress. Gomez and students had previously created the algorithm Deep Mapper, which compares a wound’s healing path with a standard path, and learns how to predict the healing trajectory of the wound and observe applied treatments’ impact on it. This allows treatment to be personalized to the subject. Furthermore, it includes a graphical user interface (GUI) for evaluation by a human physician, if necessary.

To test the effectiveness of a-Heal and its therapies, the team observed its effects on a 20mm porcine (pig) wound over the course of 22 days, as porcine and human wounds share many similarities. Two individual animal experiments were conducted, with two wounds treated with a-Heal and four as the control group. The animals did not undergo a change in routine in addition to the treatment. a-Heal was applied for the first seven days of the experiment, as previous studies displayed this to be the most optimal for healing. Both EF and Flx therapy were used during the experiment, with EF used first to guide the probability of inflammation of the wound to 40%, at which point a-Heal switched to administer Flx to advance the healing into the proliferative phase, where new tissue is formed at the wound site, with the dosage varying per the ML Physician’s instructions. As an excess of Flx can instead hinder wound progress, the ML physician is only able to prescribe a finite amount.

The results indicated that Flx has a strong positive effect on the wound healing process, especially coupled with EF, already known to be conducive towards wound healing. The team will continue to run experiments to observe the ML Physician’s adaptability and personalization of care. For now, a-Heal proposes an efficient method of wound treatment independent of hospital visits. Further merging of machine-learning technology and medicine is quite tangible.