More than three quarters of large companies today have a “data-hungry” AI initiative under way — projects involving neural networks or deep-learning systems trained on huge repositories of data.
Yet, many of the most valuable data sets in organizations are quite small: Think kilobytes or megabytes rather than exabytes. Because this data lacks the volume and velocity of big data, it’s often overlooked, languishing in PCs and functional databases and unconnected to enterprise-wide IT innovation initiatives.
But as a recent experiment we conducted with medical coders demonstrates, emerging AI tools and techniques, coupled with careful attention to human factors, are opening new possibilities to train AI with small data and transform processes.
For every big data set (with one billion columns and rows) fueling an AI or advanced analytics initiative, a typical large organization may have a thousand small data sets that go unused. Examples abound: marketing surveys of new customer segments, meeting minutes, spreadsheets with less than 1,000 columns and rows. In our experiment, it was annotations added to medical charts by a team of medical coders — just tens of annotations on each of several thousands of charts.
Medical coders analyze individual patient charts and translate complex information about diagnoses, treatments, medications, and more into alphanumeric codes. These codes are submitted to billing systems and health insurers for payment and reimbursement and play a critical role in patient care.
Coders in our experiment, all of whom were registered nurses, were already accustomed to drawing on an AI system for assistance. The AI scanned charts and identified links between medical conditions and treatments and suggested the proper code for a given chart.
We wanted to see whether it was possible to transform the coders, responsible for the accurate, one-at-a-time assessment of charts, into AI trainers capable of enriching the AI with medical knowledge that would improve the system’s performance at identifying links.
What we learned over the course of the 12-week experiment is that creating and transforming work processes through a combination of small data and AI requires close attention to human factors. We believe that three human-centered principles that emerged from the experiment can help organizations get started on their own small data initiatives:
Balance machine learning with human domain expertise. A number of AI tools have been developed for training AI with small data. For example, few-shot learning teaches AIs to identify object categories (faces, cats, motorcycles) based on only one or a few examples instead of hundreds of thousands of images. In zero-shot learning, the AI is able to accurately predict the label for an image or object that was not present in the machine’s training data. In other words, it can correctly identify things it has never seen before. Transfer learning involves transferring knowledge gained from one task to the learning of new tasks — for example, identifying subtypes of cancer, based on knowledge of another type — which eliminates the machine’s need for a vast set of new data for performing the new task.
In our experiment, we employed a tool commonly called a knowledge graph, which explicitly represents the various relationships between different types of entities: “Drug A treats condition B,” “Treatment X alleviates symptom Y,” “Symptom Y is associated with condition B,” etc. It succinctly captures expert knowledge and makes that knowledge amenable to machine reasoning — for example, about the likelihood of a specific condition being present given the drugs and treatments prescribed.
To enable the coders to impart their knowledge to the AI, we developed an easy-to-use interface that allowed them to review contested links in the graph’s database. These were links where their colleagues, when reviewing individual charts, had disagreed with the AI — either by adding links unknown to the system, or by removing links it had added. Based on their expertise, the coders could directly validate, delete, or add links and provide a rationale for their decisions, which would later be visible to their coding colleagues. In addition, they were encouraged to follow their inclination to use Google (often with WebMD) to research drug-disease links, going beyond what they regarded as the existing AI’s slow look-up tool.
This combination of machine learning and human expertise has a significant multiplier effect. Instead of merely assessing single charts, coders added medical knowledge that affects all future charts. Further, with the AI taking on the bulk of the routine work, the need for screening of entire medical charts is greatly reduced, freeing coders to focus on particularly problematical cases. Meanwhile, data scientists are freed from the tedious, low-value work of cleansing, normalizing, and wrangling data.
Focus on the quality of human input, not the quantity of machine output. In the existing system, coders focused on the assessment of individual charts in high quantity. Over time, the AI learned from the accumulation of links added or rejected by a multitude of coders: Once a drug-disease link that the AI was not familiar with had been proposed a significant number of times by coders, a data scientist added it to the graph database. This manual process was undertaken only occasionally, in part because of the time lag in accumulating link proposals, and it relied on quantitative support for the link, rather than on medical expertise.
In the new system, coders were encouraged to focus less on volume of individual links and more on instructing the AI on how to handle a given drug-disease link in general, providing research when required. Links could now be considered for addition to the knowledge graph AI with a lesser burden of quantitative evidence. The AI would learn more regularly and dynamically, especially about rare, contested, or new drug-disease links.
Recognize the social dynamics in play on teams working with small data. In their new roles, the coders quickly came to see themselves not just as teachers of the AI, but as teachers of their fellow coders. Most importantly, they saw that their reputations with other members of the team would rest on their ability to provide solid rationales for their decisions. They spoke often of the importance of those rationales to the confidence of a subsequent coder encountering an unfamiliar link.
After only a few experimental sessions, a number of the participants asked that the number of characters in the tool’s rationale textbox be increased. Later, they asked that the research box be altered to accommodate more than one reference. Notably, they not only began to devote more time to each case than they had with the existing system, but to provide even more comprehensive rationales for their decisions as the experiment unfolded. Moreover, coders indicated they felt more satisfied and productive when executing the new tasks, using more of their knowledge, and acquiring new skills to help build their expertise. They also felt more positive about working with AI on a daily basis.
As small-data techniques advance, their increased efficiency, accuracy, and transparency will increasingly be put to work across industries and business functions. Think drug discovery, industrial image retrieval, the design of new consumer products, and the detection of defective factory machine parts, and much more.
But competitive advantage will come not from automation, but from the human factor. For example, as AI plays an increasingly bigger role in employee skills training, its ability to learn from smaller datasets will enable expert employees to embed their expertise in the training systems, continually improving them and efficiently transferring their skills to other workers. People who are not data scientists could be transformed into AI trainers, like our coders, enabling companies to apply and scale the vast reserves of untapped expertise unique to their organizations. Further, the results that emerge from small-data applications will come not from a black box, as they do in data-hungry applications, but from human-machine collaboration that renders those results explainable and therefore more trustworthy both inside and outside the organization.
Mastering the human dimensions of marrying small data and AI could help make the competitive difference for many organizations, especially those finding themselves in a big-data arms race they’re unlikely to win.
By H. James Wilson and Paul R. Daugherty
Why hasn’t artificial intelligence fully transformed supply chains? Several years ago, some of us predicted that AI-powered automation would lead to “the death of supply chain management.” However, despite heavy investments, companies have not realized the vision of AI-managed supply chains.
According to our survey, only 22% of workers globally rank compensation as the thing that matters most to them in a job. This isn’t to say that people will accept a job without fair pay: Compensation still ranks higher than all other job attributes. But it’s evident that a coin-operated view of workers, where firm leaders see employment as a purely financial transaction, underestimates the deeper human motivations for work.
In November 2019 Stanford Health Care moved into a new hospital building. With seven stories and 824,000 square feet, the hospital required over a decade and two billion dollars to plan and construct. Most descriptions of the hospital focus on the airy private patient rooms or the state-of-the-art operating rooms, but one of the most technologically sophisticated aspects of the building is found in the basement.