59.6% of BH-flagged responses map to at least one PHQ-9, GAD-7, or PROMIS Global Health item.
Think i mentioned this before, but really curious about the other 40% that do not map onto these instruments. What are these responses about? Would be helpful to know in the case that we do develop a structured data collection process.
1 Research Motivation
I might include an Executive Summary at the top, which describes at a high level the intro, methods, results, and recommendations
The five patients below are drawn from, and are the strongest examples within, this smaller repeat-CSAT group.
This is such a cool way to display the qualitative data!
Below is a single ordinal composite that collapses valence, temporality, and impact_specificity into an “improvement arc”,
I might define the three variables (valence, temporality, and impact_specificity) in a bit more detail!
This section asks a different question: is there evidence of improvement over time? The Improvement Arc composite below looks for a before / after narrative within a single response, a patient’s own retrospective account of change, abundant across the sample (n ≈ 1,009) but not, strictly speaking, two separate observations of the same person.
Think this points to the need for structured data collection as well, which can help standardize future analyses.
% mental-health dx
Curious whether there are more granular dx codes for these MH conditions to map onto, similar to how you map onto the specific SDOH z codes below
the plots that follow investigate the patterns of co-occuring constructs
Very cool way to illustrate!
These are rows with genuine BH signal that didn’t map cleanly to any specific construct, a residual worth keeping in mind when reading construct-level prevalence as a share of “all BH signal.”
I'd be curious to know what types of responses fall under the 10.1%!
Each flagged response is scored against eight named constructs spanning depression, anxiety, overwhelm, coping and functioning, hope / hopelessness, loneliness / isolation, caregiver burden, and general distress.
Sorry if i missed this, but how are constructs defined? based on a theoretical framework, LLM categorization, or something else? Might be explicit here
Across these dimensions, the impact evidence above skews clearly positive.
We don't create some sort of composite score within "impact evidence", right? ie., a response coded as relief improvement and attributed to solace is scored greater than a response coded as distress and vague attribution to solace?
I guess it might be hard to do that since some aren't necessarily ordinal. Just trying to think about some ways to further quantify "impact"
For one, patients who submit CSATs are very likely unrepresentative of the entire population of Solace patients, particularly those who experience severe symptions of behavioral health challenges.
Yes- I think this is an important point to consider as we think about the type of impact we can make on those who would need it the MOST. There's obviously a response bias here. This points well to your recommendation that we should find a way to systematically collect more BH info from patients in a less-burdensome way
All 13,512 positive CSATs (from 8,853 unique patients) where scored in an initial pass to identify BH signal.
Could you look at negative csats? do we have these? curious whether this might paint a more rounded picture of Solace’s impact, instead of solely relying on the positive csats