Every time someone mentions how machines and artificial intelligence are threatening to take away the jobs of human workers, I’m reminded of the American folktale of John Henry. Hammer in hand, the legendary steel driver outpaces a newfangled steam-powered rock drilling machine, proving that human pride and hard work will outrace the progress of soulless technology. But it’s a hollow victory. In the tale, Henry dies of a heart attack—or, more likely, silicosis—soon after the drilling machine breaks down.
There’s a lesson here that is a step beyond the proverbial man vs. machine. It’s less a tale of heroism, and more a lesson on the folly of failing to recognize the value of this man. We can repair a machine, but we cannot so easily replace a skilled worker. Why not simply teach Henry to use this new tool? A man of his experience would no doubt know where to place a chisel to best effect, no matter the mechanism that drives it home. But instead, by pitting them against one another, we lose man and machine alike.
Machine, Meet Man
It is then no surprise that researchers are calling for medical device manufacturers to take a closer look at the human factor. With designs that account for human capabilities and limitations alike, modern medical devices can avoid the fate of the Henry legend’s steam-powered rock driller. Led by Penelope M. Sanderson, Ph.D., professor of cognitive engineering and human factors at The University of Queensland, an international team of researchers recently showed how clinicians have a hard time interpreting the tones emitted by pulse oximeters—a technology that has been in use for decades to indicate a rise or fall in how much oxygen a patient’s blood is carrying (SpO2). The work was published in BI&T, the peer-reviewed journal of health technology and sterilization from AAMI.
Subscribe to the MedTech Intelligence weekly newsletter to stay up to date on the latest news and information on MedTech design and development.
“The variable-pitch tone is an appreciated and trusted feature of the pulse oximeter’s user interface. However, studies show that it supports relative judgments of SpO2 trends over time and is less effective at supporting absolute judgments about the SpO2 number or conveying when SpO2 moves into clinically important ranges,” Sanderson and her team explained.
Citing studies reaching back to when the pulse oximeter was first introduced in the 1990s, the researchers showcased how the otherwise-accurate technology is limited by the ability of a user to differentiate the warning tones of the device. These studies showed that clinicians “systematically underestimate degrees of desaturation below 90% SpO2” and a worrying number of study participants were unable to identify the direction of change for two tones one step apart.
Fortunately, the researchers note that there are ways to improve the situation. Multisensory training, for instance, has been shown to marginally improve a clinician’s ability to detect changes in the device’s tone even with ambient noise, such as the cacophony of your average hospital. However, most promising may be the authors’ own suggestion: Rather than smooth incremental changes across all SpO2 values, the next generation of pulse oximeters could indicate whether SpO2 is above or below target values, and to what degree, using specific acoustic signatures, such as changes in timbre and tremolo.
“We envision that these ranges would be adjustable by clinicians, who would tailor them for individual patients just as they do for alarm thresholds,” the team wrote. “Acoustic enhancements are effective in the absence of alarms, potentially reducing contributions to alarm fatigue, but alarms could still sound at selected levels if and as required. Acoustic features such as beacons, changes in timbre, tremolo, or formants could be used as enhancements, and they yield substantial benefits compared with other approaches, such as manipulating pitch step sizes or specialized training.”
Machine, Meet … Machine?
But what about using another machine to help man and machine better work in tandem? In a new BI&T research article, engineers and scientists from Purdue University showcased how human-machine learning (ML) collaboration provides more insights into the administration of drugs through smart infusion pumps than human monitoring and data logging alone, resulting in better patient safety.
“Avoidable medication errors related to the use of intravenous (IV) medications occur more frequently than other types of medication errors, with frequencies between 48% and 81%,” the authors reported. They explained that, while dose error reduction software (DERS) already exists to prevent caretakers from mistakenly administering drugs beyond safe dosage limits, a preset drug library “might not necessarily reflect varying patient needs.” The limits of this technology also do not account for the insights of clinicians on the front line, calling for them to override the software when necessary. Reminiscent of the John Henry folktale, this sets doctors against the technology that’s supposed to improve their profession.
Fortunately, the implementation of additional technologies, not less, may help humans better inform DERS. The authors describe how the data collected by automated infusion pumps and their DERS can be fed into three different specialized ML algorithms, allowing for regular analysis and the identification of potentially dangerous anomalies in patient care.
In nearly 175,000 instances of delivering propofol to patients through infusion pumps over 365 days, 3,300 alerts were triggered by DERS monitoring. The ML algorithms identified 31 unique anomalies in this data, calling for closer human examination. A chart created through manual human data analysis alone only identified 15 anomalies.
“All three ML algorithms agreed that an unusual infusion alerting pattern occurred on July 30, while the [chart] found no anomaly on that day,” the authors provided as an example. “Although the maximum times-limit value on that day was 3 (below the fatal threshold), 71 alerts were generated at an average rate of 1.06 per unique infusion. One-half of those alerts were also overridden [by clinicians] … This hints at the need to go beyond control charts to identify unusual alerting patterns.”
They added, however, that the human-generated charts were easier for clinicians to interpret, “lending credence to the importance of combining multiple relevant variables for anomaly detection.”
“Our results showed that beyond reviewing pump reports, plotting aggregated univariate charts, and applying basic statistical methods, a mix of intuition, domain expertise, computational rigor, and human-ML collaboration provides more insights into the infusion administration process,” they concluded. “A more effective means of identifying unusual infusion alert patterns can be achieved as a first step to advance safer IV medication administration.”
- M. Sanderson et al (2022) “Signaling Patient Oxygen Desaturation with Enhanced Pulse Oximetry Tones.” BI&T. https://doi.org/10.2345/0899-8205-56.2.46
- Obusen et al (2022) “Detecting Unusual Intravenous Infusion Alerting Patterns with Machine Learning Algorithms.” BI&T. https://doi.org/10.2345/0899-8205-56.2.58