Evidotes: Evidence & Emotional Support
- Evidotes is an information support system that integrates scientific evidence and human anecdotes to address uncertainties in online peer health posts.
- It employs three distinct lenses—Dive Deeper, Focus on Positivity, and Big Picture—to tailor information retrieval and regulate emotional responses.
- A pilot study with chronic illness patients demonstrated improved information satisfaction and reduced emotional costs, highlighting its practical impact.
Evidotes is an information support system for online peer health posts that treats such posts as uncertainty triggers and augments them, in place, with a curated blend of scientific evidence and human anecdotes. Implemented as a Chrome browser extension for Reddit, it retrieves relevant material from PubMed and Reddit, synthesizes claim-level takeaways, and presents them through three user-selectable lenses—Dive Deeper, Focus on Positivity, and Big Picture—each corresponding to a different uncertainty-management strategy. Its central premise is that the informational and emotional consequences of reading peer experiences are not adequately addressed by systems focused only on relevance, credibility, or fact-checking; instead, readers often need contextualization, emotional regulation, and structured ways to navigate uncertainty. In a mixed-methods study with 17 chronic illness patients, Evidotes improved self-reported information satisfaction from 3.2 to 4.6 and reduced self-reported emotional cost from 3.4 to 1.9 relative to baseline browsing (Bali et al., 18 Feb 2026).
1. Problem framing and conceptual basis
Evidotes was designed for a setting in which people managing chronic illness use online communities to learn from lived experience, gather practical tips, and find empathy. In this setting, peer health posts are valuable precisely because they are vivid and personally resonant, but those same properties can provoke new uncertainties: whether a side effect could happen to oneself, whether alternatives exist, or whether a reported approach would work for someone similar. The system therefore frames peer posts not primarily as truth claims to be screened, but as triggers for emergent information needs and emotional responses (Bali et al., 18 Feb 2026).
This framing differs from accuracy- and relevance-centered systems that emphasize credibility assessment, fact-checking, and filtering. According to the paper, such approaches do not directly address three recurrent problems: emergent information needs that arise during reading, emotional costs such as anxiety and discouragement during unstructured browsing, and an “integration paradox” in which readers say they want scientific sources but remain in forum threads because of search friction. Evidotes centers the need to contextualize stories, regulate emotional exposure, and actively navigate uncertainty rather than spiraling (Bali et al., 18 Feb 2026).
The conceptual contribution is therefore not only technical retrieval or summarization. It is an uncertainty-management model in which information augmentation is tuned to moment-to-moment needs. Sometimes a reader wants technical specificity, sometimes hope and positive trajectories, and sometimes perspective and broader management strategies. This suggests a shift from static notions of “better information” toward adaptive support for different epistemic and affective states.
2. Lens-based system design and interface structure
Evidotes operationalizes uncertainty management through three explicit lenses. A Chrome extension injects an “Evidotes” button next to Reddit post titles; selecting a lens opens a draggable, minimizable overlay panel at the bottom-right. The panel contains two tabs—“Insights for You” and “Profile”—and the synthesis view begins with framing text that transparently states the lens’s intent. Claims are then shown as short takeaways with expandable sections for “Research Findings” and “User Anecdotes,” each linked and quoted for verification (Bali et al., 18 Feb 2026).
| Lens | Query emphasis | Intended use |
|---|---|---|
| Dive Deeper | condition, medication, side effect, timelines | information seeking |
| Focus on Positivity | success, recovery, improvement, positive outcomes | emotion regulation |
| Big Picture | management strategies, guidelines, alternatives | cognitive reappraisal and perspective broadening |
Dive Deeper is analytical and emphasizes entities and specifics such as side-effect likelihoods, timelines, mechanisms, and alternatives. Focus on Positivity biases retrieval toward “success,” “recovery,” “improvement,” and “managed well,” and its synthesis tone is hopeful; no sentiment classifier is used, so positivity is controlled through query bias and synthesis constraints. Big Picture broadens the frame toward guidelines, lifestyle modifications, dietary changes, management strategies, and complementary approaches, with a neutral and strategic tone (Bali et al., 18 Feb 2026).
The implementation stack is correspondingly lightweight but structured. The frontend is a Chrome extension using Vue.js, the backend is a Flask service with Agno-AI orchestration, and the LLM is GPT-4o-mini with temperature 0. Retrieval uses PubMed via Entrez.esearch and Reddit via PRAW and the Reddit Search API, without custom indexing; latency is managed through asynchronous requests and a 25-second timeout (Bali et al., 18 Feb 2026).
3. Retrieval, synthesis, and validation pipeline
The retrieval pipeline begins by extracting health entities from a post—condition, treatment, and symptoms—and combining them with lens-specific intents to form multiple search strings. Parallel search across PubMed and Reddit retrieves approximately 10 results from each source. PubMed results are accepted in PubMed Best Match ranking, which weights MeSH term matching, title and abstract keyword frequency, and citation signals; Reddit results are taken in API “relevance” order, which reflects query match, post score, and engagement. No additional pre-filtering is applied, because the LLM assesses relevance from full context during synthesis (Bali et al., 18 Feb 2026).
Synthesis is lens-specific retrieval-augmented generation. The prompt enforces the intended style—analytical, hopeful, or neutral—requires quotes for each claim, and prioritizes claims supported by both evidence types. The panel preserves source transparency by showing expandable quotations and links for verification; when one source type is absent, the corresponding section remains empty rather than being backfilled. If nothing relevant is found, Evidotes returns empty arrays, explicitly avoiding hallucinated claims (Bali et al., 18 Feb 2026).
Post-hoc validation was designed around citation faithfulness, medical accuracy, and topical relevance. Faithful citation checks used spaCy normalization and BERTScore with a threshold of at least 0.85; 96.7% of quote-source pairs matched their sources, with 77.3% verbatim and 19.5% high-similarity paraphrase. At the claim level, 89% of claims were fully supported by all cited quotes, 11% were partially supported, and only three were unsupported. A clinician judged 69 of 70 sampled claims medically accurate, with none harmful. BERTopic was used post hoc to assess topical relevance, yielding 98.2% overall relevance, with Dive Deeper and Focus on Positivity at 99.4% and Big Picture at 94.7%; the paper notes that the lower Big Picture figure is expected given broader retrieval (Bali et al., 18 Feb 2026).
These design choices place Evidotes in a distinctive position among RAG systems. Rather than optimizing only relevance or factuality, it explicitly encodes support gaps, source asymmetries, and lens-conditioned synthesis constraints.
4. Co-presentation and “information symbiosis”
A core concept in Evidotes is “information symbiosis,” defined as the mutual reinforcement of scientific evidence and anecdotes when they are co-presented rather than arranged in a credibility hierarchy. Anecdotes function as gateways to research: participants often read Reddit first, and side-by-side display made it effortless to jump into relevant PubMed articles that they otherwise would not seek. Research, conversely, functioned as a filter and generalizer: PubMed context helped participants assess which stories were credible or useful and extract population-level takeaways without being swayed by extreme cases (Bali et al., 18 Feb 2026).
The interface realizes this symbiosis at the claim level. A single claim can display a research section with PubMed citations, quotes, and paraphrases, next to a user-anecdote section containing Reddit links and quotations. The reader can switch lenses on the same post, expand and collapse claims, skim quotes without leaving the thread, and click through to original sources. Empty sections are preserved when only one source type supports a claim, making incompleteness visible rather than concealed (Bali et al., 18 Feb 2026).
The paper’s illustrative PCOS example makes the mechanism concrete. For a post about severe Metformin side effects, Dive Deeper can surface claims about side-effect prevalence, typical timelines, and alternative medications; Focus on Positivity can foreground success stories and improvements involving Metformin or alternatives; Big Picture can foreground broader PCOS management strategies such as diet, exercise, other treatments, and guidelines. In this framing, research makes peer tips less overwhelming, while peer tips make research actionable in everyday life (Bali et al., 18 Feb 2026).
A plausible implication is that Evidotes treats evidence integration as a usability and sensemaking problem rather than merely a ranking problem. The side-by-side structure reduces search friction, preserves provenance, and lets the same post be re-read under different cognitive and emotional frames.
5. Empirical study and observed usage patterns
The evaluation used a mixed-methods protocol with 17 adults aged 18 to 64 who had diagnosed chronic conditions across seven ICD-11 categories, including nervous system diseases such as migraines, endocrine and metabolic diseases such as diabetes, cancer, and musculoskeletal disease. The procedure consisted of a baseline survey and interview, tool introduction, approximately 25 minutes of open exploration of condition-specific subreddits using Evidotes, and a post-session survey and interview. Measures included 5-point Likert scales for information satisfaction, mental demand, and stress or annoyance, along with items on whether the tool delivered wanted information and reduced emotional overwhelm (Bali et al., 18 Feb 2026).
Quantitatively, information satisfaction improved from 3.2 to 4.6, a 43% change with and Cohen’s . Stress or annoyance decreased from 3.4 to 1.9, a 44% change with and . Fifteen of 17 agreed that the tool gave the information they wanted, and 16 of 17 agreed that it reduced emotional overwhelm. The study reports the standard definition of Cohen’s but does not specify confidence intervals or test type (Bali et al., 18 Feb 2026).
Usage logs indicate structured but flexible interaction. Participants browsed 77 posts, made 130 lens interactions, and received 293 synthesized claims, with a mean of 2.25 claims per post and standard deviation 1.20. Each claim had on average 2.12 sources, approximately one PubMed source and one Reddit source. Dive Deeper dominated initial use, appearing on 61 of 77 posts, with common transitions to Positivity and Big Picture. Condition category influenced lens preference, with and ; neoplasms favored Positivity, while migraines and diabetes favored Dive Deeper. Cluster analysis identified a majority of “diverse navigators,” who adjusted lenses by post, and a minority of “focused deep-divers,” who consistently used Dive Deeper (Bali et al., 18 Feb 2026).
Qualitatively, the paper reports three themes: source symbiosis, explicit lenses as a means of reducing overwhelm and teaching new strategies, and bounded exploration in which the overlay panel served as an anchor that balanced curiosity with focus. This suggests that the observed gains were not only due to retrieval quality, but also to interaction design and framing.
6. Reliability, limitations, ethics, and broader evidence-system context
Evidotes incorporates several safeguards intended to make augmentation transparent and medically non-harmful. Scientific evidence is sourced from PubMed and experiential knowledge from Reddit; source types remain explicit and comparable. Citation faithfulness checks, URL accessibility checks, and expert review were used as post-hoc reliability mechanisms. At the same time, the paper identifies residual risks: convenient workflows may discourage deep source verification, the Positivity lens could bias exposure over time, and a small number of claims reflected tenuous inference, motivating stricter future synthesis constraints (Bali et al., 18 Feb 2026).
The reported limitations are substantive. The study is small (), single-session, and Reddit-specific; image-heavy posts are not supported. Anecdote selection inherits platform biases, including who posts and whether posting skews negative or positive. Evidotes does not require health disclosures for personalization, and the study used consent, recorded sessions, and self-report measures, but the paper does not present long-term behavioral evidence. Proposed future guardrails include progressive disclosure, verification prompts, pattern-based reminders such as encouraging a switch from Positivity to Dive Deeper, hallucination identifiers, and usage analytics to detect sustained one-lens use (Bali et al., 18 Feb 2026).
The broader significance of Evidotes becomes clearer when placed alongside other evidence-centric NLP systems. “Evidence Inference 2.0” models comparative-effect inference in RCT articles and evidence span extraction, reaching macro-F1 0.780 with a Biomed RoBERTa pipeline (DeYoung et al., 2020). “TwoWingOS” jointly optimizes evidence selection and claim verification, achieving ScoreEv 54.33 and Evidence F1 47.15 on FEVER (Yin et al., 2018). “u-EIDG” mines evidence veracity labels and applies evidence-focused attention in knowledge-intensive dialogue generation, improving MultiDoc2Dial performance to 46.85 token F1, 33.24 SacreBLEU, and 44.65 ROUGE-L (Wu et al., 2023). A plausible implication is that Evidotes belongs to the same general family of evidence-aware systems, but differs in its target outcome: uncertainty navigation under emotional load rather than only entailment accuracy, trial-result extraction, or dialogue factuality.
Within that broader landscape, Evidotes reframes health information support as the curation of evidence and anecdotes around a reader’s immediate uncertainty state. Its distinguishing claim is not that evidence should replace anecdote, or vice versa, but that the two become more useful when structured, co-presented, and filtered through explicit lenses of inquiry, positivity, and strategic perspective.
