Varied Uses for Artificial Intelligence in Health Care Show AI’s Riches

Industries, researchers, and governments have only scratched the surface of what they can do with artificial intelligence. Recently I spoke with three companies using AI in creative and unexpected ways in health care. Here are their stories.

Personalized Medication Administration by Dosis

Over-the-counter cough medicines and aspirins make dosage sound simple: a teaspoon of liquid or two tablets every four hours. Most people don’t realize that finding the right doses for many medications is very difficult, with proper dosage varying a good deal from one person to another. According to Shivrat Chhabra, CEO of Dosis, some medications, such as erythropoietin stimulating agents (ESAs) needed by people with chronic anemia, can vary by a multiple of 30.

Figure 1 shows a chart that compares the doses and hemoglobin levels for a patient with anemia.

A graph shows doses and hemoglobin levels, followed by tables of information.
Figure 1: Doses and hemoglobin levels

Dosis determines the right dose for each individual iteratively. When a patient first starts a medication, Dosis recommends an initiation dose. Over time, Dosis uses data on the medication administered and the responses seen in laboratory results to more precisely model the individual’s response to the drug and recommend optimal dosing. Patients with anemia, for instance, have a laboratory test to check their hemoglobin level periodically–typically every two weeks or once a month–which results in a dose adjustment. Dosis has found that it can help reduce the amount of drug required to effectively manage anemia by an average of 25 percent.

Chhabra reports that Dosis modeling takes a burden off of doctors, allowing them to focus additional time and attention on patients that require the most attention.

Recognizing Suicidality Through an AI-Based Voice Assistant by MyndYou

I remember decades ago when you had to train voice recognition software for weeks to recognize your voice. Modern voice-activated devices work out of the box. But the horizons for voice recognition are expanding, with a demand for lots more research.

In health care, sophisticated voice recognition could be a matter of life and death, as we’ll see in the following discussion of an AI-based voice assistant provided by Cosán Group, based on software from MyndYou.

Cosán Group provides chronic care management, remote patient monitoring, and behavioral health integration for aging people. They rely on automated interactions with clients for much monitoring. The particular MyndYou assistant discussed in this article is trained to recognize suicidality. It interprets not only what that the patient is saying but their tone of voice.

I exchanged email with the MyndYou team to explore the many subtleties and unusual requirements of such an application. The health care field has recently started to come to grips with disparities in care for people of different genders, races, and ethnicities. MyndYou trained its machine learning model on input from a wide variety of patients. Their software recognizes both English and Spanish, as well as multiple dialects.

The first release of the assistant was based on tens of thousands of hours of call recordings between patients and clinicians. The assistant could learn from nurses and therapists with decades of experience conducting high-quality interviews. Thus, the assistant has achieved the capability of responding sympathetically and warmly, not just mechanically posing questions.

An important part of their development process consists of ethical case reviews, conducted both within the company and with their clinical partners. The company also reviews the questions posed by the assistant to make sure everyone will find them clear. The questions are also designed with the personal attributes of the client in mind: age, gender, culture, etc.

Unsupervised Deep Learning in Pharmaceutical Development by Neurala

The health care industry keeps uncovering new areas where it can exploit artificial intelligence (AI). In a sign that should alert manufacturers in all industries—not just health care—to the value of AI, Neurala has entered pharmaceutical development and manufacturing. The company offers its deep learning software to manufacturers producing devices used in drug development, as well as to pharma companies directly.

Recently I talked to Max Versace, CEO and co-founder of Neurala. Founded in 2006 at Boston University and still located in Boston, they have shipped their AI software in 60 million devices ranging from cameras to robots. They have collaborated with NASA and other agencies. Their strategic entry into the pharma industry took place in 2019, with the invention of a radically simplified AI.

When deep learning first started to prove its value some 25 years ago, it required both sophisticated programming skills and a doctoral-level knowledge of statistics. Programmers had to judge matters such as how many categories to group data into, and had to choose half a dozen “hyperparameters” precisely to ensure accurate results.

In recent years, though, experts at automation have figured out how to automate those tasks as well. Often, the solution is a combination of brute force and careful selection: run hundreds of algorithms and refine the results until they’re what you want. What this means for business is access to deep learning without having to woo data scientists with salaries in the hundreds of thousands of dollars.

Neurala joined this trend in 2020 with its vision inspection automation (VIA) software, which allows a customer without knowledge of AI to build visual models quickly from a dashboard. Pharma is one of the beneficiaries of the new software.

Versace described VIA as an “artificial eye” in a production line or screening environment, replacing human inspectors. A brief video gives a sense of the service. Although Neurala’s tools are useful in a wide variety of situations, one of the easiest to understand is discovering defects in vials containing medicine. Cracks, missing lids, and contaminants are recognized.

Each manufacturing environment is unique, so each client has to train a unique deep learning model. That’s why it’s so important to simplify model training and remove the need for programming or choosing hyperparameters. Neurala does provide clients with guidelines for making and selecting images properly. Clients can thus avoid everyday errors such as confusing the deep learning system with irrelevant colors.

Neurala claims that they don’t need thousands of test cases for classification, as classic deep learning does. Furthermore, a pharma company doesn’t have to divide tests into good and bad categories, as in most deep learning classification. Just provide some images of good vials, and the software will flag vials that deviate from the norm. Neurala’s software also provides explainability (sometimes called accountability), by showing the exact pixels that let the model isolate a defect.

Neurala’s VIA is also used for other pharma tasks, including compound screening, which helps a company identify more quickly the molecule that will accomplish a task. Like other applications of AI, when done with a respect for its proper role, these applications should remove tedious tasks from humans and allow them to play more creative roles in their industries.

About the author

Andy Oram

Andy is a writer and editor in the computer field. His editorial projects have ranged from a legal guide covering intellectual property to a graphic novel about teenage hackers. A correspondent for Healthcare IT Today, Andy also writes often on policy issues related to the Internet and on trends affecting technical innovation and its effects on society. Print publications where his work has appeared include The Economist, Communications of the ACM, Copyright World, the Journal of Information Technology & Politics, Vanguardia Dossier, and Internet Law and Business. Conferences where he has presented talks include O'Reilly's Open Source Convention, FISL (Brazil), FOSDEM (Brussels), DebConf, and LibrePlanet. Andy participates in the Association for Computing Machinery's policy organization, named USTPC, and is on the editorial board of the Linux Professional Institute.

   

Categories