Invest in Wellness

5,907 Messages. One Signal.

OpenAI says 230 million people discuss health with ChatGPT. That proves the interface works.

But for a payer, the question isn't whether they'll talk. The question is whether the conversation predicts the cost.

We ran an 8-week pilot with 11 seniors to find out.

The "Kara" Signal

Kara joined the pilot feeling stuck.

  • Week 2: She texted about job interview anxiety.
  • Week 4: She told us she was depressed after not getting the role.
  • Week 6: "I've given up on finding a job... I'm 64 years old."

Her sentiment scores tracked the decline in real-time. We saw the trajectory bending downward before she ever used the word "depressed."

But here's what claims data would miss entirely: Despite the emotional crash, Kara kept engaging. She kept doing yoga. Her self-efficacy (confidence in managing health) held steady.

The sentiment analysis told us she was struggling. The sustained engagement told us she was coping. Both signals mattered. Neither appears in a claims file.

The Signal in the Noise

This was a feasibility pilot designed to answer one question: Is there signal worth capturing?

We analyzed 5,907 messages over 8 weeks, tracking sentiment trajectories using VADER. This tool handles negation and context - critical for parsing casual health language (e.g., "not bad" vs. "bad"). We then measured the correlation against self-efficacy scores at Week 8.

The results were stark:

  1. Sentiment predicts outcome (r=0.63, p=0.07). Members whose language grew more positive showed greater improvement in health confidence. For a small sample (n=9 completers), a correlation this strong suggests we are measuring something real.
  2. Quality beats quantity. Message volume showed no correlation with outcomes. Our highest-volume user actually declined due to hand pain. The intervention worked; the modality didn't.

Why This Matters

Healthcare runs on lagging indicators.

  • Claims show what happened last month.
  • Labs show what broke last week.
  • Sentiment shows who is crashing right now.

We found a leading indicator buried in daily conversation. While incumbents wait for claims to trigger care management, we act on the conversation.

What's Next

We have identified the signal. The next phase scales the study and adds agentic capabilities. The AI won't just detect the decline - it will initiate the intervention to stop it.

We hear it first. We act first. We capture the savings.