Building a Burnout Metric with HR Analytics

How do you build a burnout metric with HR analytics? The key lies not in the technology but in the design principles. A good burnout metric is not meant to surveil individuals, but to create organization-level signals that cannot identify any individual. The goal of understanding depletion faster with data and the goal of protecting members’ trust may seem to conflict, but if designed together from the start, they can coexist.

01. What Is People Analytics?

People analytics (or HR analytics) is an approach that seeks to back HR-related decisions with data rather than intuition.

Traditionally, HR decisions have relied heavily on experience and gut feeling. Who is at high risk of leaving, or which team’s mood is souring, usually depended on a manager’s impressions.

People analytics is an attempt to add data to these judgments. It tries to understand topics such as hiring, attrition, engagement, performance, and well-being and burnout through data.

Behind this trend is also a broader concept called digital phenotyping. It is an approach that seeks to read signals of behavior and state from the digital traces people leave in daily life, and it is being actively discussed in mental health research. Designing a burnout metric can be seen as one application of this trend.

02. What Signals Go into a Burnout Metric?

There are several signals whose use can be considered when designing a burnout metric.

Survey. The most standard method. Examples include job-stress surveys, engagement surveys, and pulse surveys. They have the advantage of using standardized questions, but they have limits: long measurement intervals, the need for responses, and the difficulty of seeing change from a single momentary score.

Work and activity patterns. Things like the distribution of working hours and rest patterns. However, such data carries an especially high risk of misinterpretation and surveillance, so it must be handled very carefully.

Engagement signals. Things such as the degree of participation in organizational activities can be referenced.

Supplementary signals such as voice. The possibility that features like speaking rate, pitch, and pauses appearing in short voice-based check-ins are related to emotion and stress is being researched.

There is a point to be remembered here without fail. No single signal can confirm burnout. That is why combining multiple signals and, above all, looking at change over time is important.

A concrete discussion of how to treat voice signals as a metric is covered in The Voice Biomarker Guide.

03. The Biggest Trap: Individual Surveillance

The most common and dangerous trap when building a burnout metric is individual surveillance.

Imagine a situation where the metric is computed at the individual level and a manager gets to look into a specific individual’s depletion score. At first glance it seems helpful, because it looks like at-risk people could be found early.

But in reality, the opposite happens.

“If my state is shown to my manager as a score, will I be able to express myself honestly?”

The moment a member feels this way, two things collapse at once.

1. Trust collapses. Members perceive the service as the company’s surveillance tool.
2. Data becomes distorted. People do not respond honestly and instead react defensively.

In other words, a metric that tries to surveil individuals destroys the reliability of the metric itself, before any question of accuracy. Data obtained from someone who feels surveilled is already contaminated.

This problem is as important as technical accuracy—or more so.

04. Designing Anonymized, Organization-Level Metrics

So what should be done? The answer is to shift the target of the metric from individuals to groups.

SYMPLE considers the following principles important.

• To individuals, it provides only an understanding of their own state and change. Change relative to the individual’s own baseline is the core.
• To organizations, it provides only aggregate-level (aggregate, de-identified) trends and risk factors that cannot identify any individual.

Concretely, it is a difference of this form.

(Avoid) “Mr. Kim’s burnout risk has increased.” (Aim for) “Recently, stress signals related to workload have been continuously rising in a particular organization.”

The latter is sufficient for an organization to find intervention points, but it does not identify any individual.

At this point, one practically important question arises. From what minimum number of people should aggregate-level data be provided in order to sufficiently lower the risk of identifying individuals? Because if the group is too small, anonymization becomes meaningless. This is one of the research questions SYMPLE is actually verifying.

05. Voice Is a Supplementary Signal, Not a Diagnosis

There is a common misconception when putting voice into a burnout metric: that “AI diagnoses burnout from your voice.”

It does not.

Voice is influenced by a very wide range of factors—sleep, colds, the surrounding environment, the microphone, gender, age, language, an individual’s speaking habits, and more. Confirming an individual’s burnout from a single voice sample requires ample caution.

That is why SYMPLE treats voice as one supplementary signal. What matters is not a single instance of the voice, but whether a meaningful change appears when compared with that person’s usual state (their individual baseline).

More detail on voice stress signals can be found in The Voice Stress Biomarker.

06. In Summary: The Standard for a Good Metric

Building a burnout metric with HR analytics is, in the end, a balance of two goals: understanding depletion faster, and protecting members’ trust.

• People analytics is an approach that adds data to HR judgments.
• A metric can include surveys, activity patterns, engagement, and supplementary signals such as voice.
• Individual surveillance is a trap that destroys both trust and data quality at once.
• Metrics should be designed at the anonymized, aggregate level.
• Voice is a supplementary signal for seeing change relative to an individual’s own baseline, not a diagnosis.

The standard for a good burnout metric is not only accuracy. Whether who can see what was designed together determines the reliability of that metric.

---

References

1. Insel TR. Digital phenotyping: technology for a new science of behavior. JAMA. 2017;318(13):1215–1216.
2. Maslach C, Schaufeli WB, Leiter MP. Job Burnout. Annual Review of Psychology. 2001;52:397–422.
3. World Health Organization. Burn-out an “occupational phenomenon”: International Classification of Diseases. 2019.
4. Maslach C, Leiter MP. Understanding the burnout experience: recent research and its implications for psychiatry. World Psychiatry. 2016;15(2):103–111.