The Human Signal in the Data Noise

Although essential, much of the data generated during healthcare delivery is difficult to categorize and/or quantify, making it nearly impossible to incorporate into data dashboards used to inform clinical workflows and healthcare delivery. A fearful patient. A hesitant clinician. A nurse who realizes that a set of discharge instructions is not being understood. These instances are not fields in a data set. They are experiences and contain information that organized data frameworks fail to capture.

AI is intended to provide a means to help capture and catalog this data. For it to be truly disruptive, it must learn to listen in addition to reading. If we only focus on measuring what is easy to categorize and quantify, we lose meaningful data that provides context for every singular clinical and healthcare delivery interaction. In the end, a model that is trained on data that is easy to categorize and quantify (data that has been stripped of meaning) will produce outcomes that are contextually equally as meaningless. The next significant step in the evolution of AI in healthcare will not focus on expanding the datasets used. It will focus on enhancing the technology in order to allow it to better "listen."

Why We Stopped Listening

Why is it that the most critical information is often the least documented? The answer has to do with mechanics. The information that is easy to measure is managed and recorded. With the absence of structure, information is discarded. Counting the number of visits is easy. Documenting a patient feeling that no one listened is not, and this feeling will never be documented. Gradually, the system optimizes for the measurable and neglects the unmeasurable. The signal did not vanish; we just built systems that are unable to measure it.

The Signal Lives in the Unstructured

In healthcare, the human signal exists in the unstructured data. In the free text of clinical notes, in a voice recording, in the messages exchanged in the care communication system, or in the comments provided by a patient in the feedback box at the end of the visit. These are the data that reveal the thoughts of a clinician or the feelings of a patient, and these are the data for which most systems were never built to capture.

Knowing the purpose of the majority of electronic records makes it easy to see the deficits. Most records capture billable events and reportable outcomes, but miss everything in between and beyond. The newer models offer deeper analysis by interpreting spoken and written language. This analysis can reveal the first signs of employee distress and patient dissatisfaction. Properly interpreted, the notes and call logs help the system finally hear what its caregivers have been hearing all along.

Experience Data

If you stop compartmentalizing the clinical, the operational, and the emotional, you begin to experience the benefits of what some call experience data. This data integrates the structured data of record systems with unstructured notes and call transcripts. When all of this data is captured and integrated, it becomes possible to view the continuum of care.

I observed this issue firsthand. A hospital I collaborated with received several complaints from patients regarding communication and assumed it was an issue of increased patient volume. However, looking closely at the complaints, we found that clinician burnout was the primary issue. Communication issues occurred in the hospital units that were the most staff depleted. This finding was not derived from a dashboard. It was derived from a careful reading of the complaints to understand what the patients were really raising.

This finding has significant practical implications. Analytic dashboards can lead to the assumption that you should add staff to handle volume, while a careful reading of the complaints shows that the real fix is addressing staff burnout. This example demonstrates that the correct solution can be derived from the same complaints, and without the qualitative reading, the staff burnout issue would continue to exist. The ability to convert qualitative data into actionable information is one of the most unused analytic tools in healthcare.

When Listening Goes Wrong

While some listening practices are better than others, candidness about the failures is important. The prime example is the almost universal tendency to collapse a situation to a number. A model reads a note, stamps it negative, rolls it into a number, and the number is easy to chart and almost meaningless, because it has taken away the one thing that mattered, the reason. A patient is dissatisfied tells you nothing. A patient is dissatisfied because no one told him why the surgery was delayed tells you exactly what to address. Chase a better score while ignoring the reason, and all you have built is a faster way to miss the point.

The second failure is quieter and more corrosive. Once staff realize their input is being scored, they begin either withholding it or writing for the scorer. Notes become more neutral, and feedback becomes more prudent, and the honest signal you were trying to capture dries up at the source. Measuring a thing can change the thing, and in this case it can erase exactly the candor that made the data worth reading. Listening only works while the people being listened to still believe it is safe to speak plainly.

Listening Ethically

Listening ethically demands discipline. Just because something is analyzable does not mean it is appropriate to analyze it. Healthcare analytics contain data that can be very personal to an individual, and insight extraction, analysis, and data interpretation is not a neutral act. It must be done with the permission of the individual or stakeholders and within context, and with a clear understanding of the purpose and whose interests the analysis serves.

Listening ethically means the analysis empowers the people in the data rather than exploiting them. It means protecting dignity while you deepen understanding, and holding any analysis of voice, speech, or text to the same standard of care you would apply to any other clinical decision. When speech and emotions are treated as data, the obligation to protect that data is the same as protecting any other personal healthcare record. The moment an analysis or interpretation of voice, speech, or text serves an organization at the expense of the individual, it has become the thing it was designed to prevent.

From Insight to Action

This should be understood as a way to expand perception rather than to replace judgment. Imagine a leadership team who, in near real time, identify early clinician note and patient message indicators of confusion and fatigue, and who can intervene before the indicators signal more serious problems. This is not analysis for the sake of analysis. This is an ability to direct considerable attention toward challenges that, at present, the system identifies only after the challenges have begun to create significant problems.

The concept that keeps this honest is simple. The model identifies the signal. A human reads and interprets the signal. The unit of action is the account rather than the score, as only a human in context can determine whether a statement requires a policy change or a discussion. In this context, AI is not the system that makes the call. It is the system that makes the call more likely to be made before the problem hardens.

When a system learns to listen, it is no longer purely reactive. It no longer waits for the break to respond. It learns to identify the early tremors instead. This is how data becomes human. It no longer counts. It learns to understand the needs of those the system was built to serve.

What are we really listening to?

The most important signal in healthcare is not a number. It is an emotional signal that is in the words people say and how they say them. It is in the words they do not say and is most important because those signals are often the most difficult to identify. Those signals often appear long before the numbers show any change.

This also shows humility. A system cannot operate on information it does not capture. A system cannot act on information it has never learned to listen to. The decision to capture the signal and to value it the same as a lab test value shows what the organization believes is important. The technology is there to make the listening possible. The choice to act on what it hears is a choice that is still up to a human.

The goal was to use technology to make the healthcare systems more human and to make the technology hear what the users are telling it. If used properly, AI can help a system hear what it otherwise could not.

Christopher Hutchins Founder and CEO, Hutchins Data Strategy Consultants