Methodology

How the AI & Mental Health Research visualization is built

Overview

This visualization maps academic papers published between 2021 and 2026 at the intersection of artificial intelligence and mental health. Papers are positioned so that semantically similar work appears nearby, colored and sized by bibliometric metadata, and organized into automatically discovered topic clusters.

Dataset collected from Semantic Scholar on February 16, 2026.

Design Approach

The visualization is built almost entirely on open-source, academically published tools from two research groups. The Allen Institute for AI provides paper discovery, metadata, and citation data via the Semantic Scholar Academic Graph API, SPECTER v2 document embeddings (which determine both the corpus boundary and the 2D map layout), TLDR summaries, and field-of-study classification. The Tutte Institute for Mathematics and Computing provides UMAP for dimensionality reduction, Toponymy for unsupervised hierarchical topic discovery, and DataMapPlot for interactive visualization.

The only proprietary component is GPT-4o, which generates human-readable names for the topic clusters discovered by Toponymy. It does not influence which papers are included, where they are placed, or how they are grouped—only what the groups are called.

This means the key decisions are algorithmic rather than editorial: the corpus boundary is set by embedding similarity to a computed centroid (not hand-curation), topic clusters emerge from unsupervised learning over the embedding space, and paper positions reflect measured semantic relationships. The goal is a map that is reproducible and resistant to selection bias.

Corpus Collection

All papers are sourced from the Semantic Scholar Academic Graph API. Building the corpus is a three-stage process: we define a center, cast a wide net, then draw a boundary.

1. Define the center

We search Semantic Scholar for “artificial intelligence mental health” using their relevance-ranked search. Unlike a simple keyword filter, this search uses a neural model that understands semantic relatedness—it finds papers that are about this topic, not just papers that happen to mention the right words. Results are limited to papers with at least 5 citations to ensure stable, well-established anchor points. The top 200 results by relevance form the anchor set—the papers most squarely at the intersection of AI and mental health.

The anchor set defines the core of the field, not its extent. Each anchor paper has a SPECTER v2 embedding—a 768-dimensional vector that captures its meaning. The trimmed mean of these embeddings (trimming 5% from each tail for robustness) becomes the centroid, the semantic center of AI + mental health research. Empirical analysis shows the centroid stabilizes by around 200 papers, while lower-ranked results drift off-topic and would dilute the center.

2. Cast a wide net

Next, we search broadly for candidate papers using Semantic Scholar's bulk search. A paper is a candidate if it mentions at least one mental health term (e.g., depression, anxiety, autism, psychotherapy, PTSD, schizophrenia, and others) and at least one AI term (e.g., machine learning, deep learning, natural language processing, large language model, and others). We use a generous list of terms on both sides—the goal here is recall, not precision. Papers are restricted to the 2021–2026 date range.

3. Draw the boundary

Finally, we use semantic embeddings to decide which candidate papers belong in the corpus. Think of it like drawing a circle on a map: the centroid marks the center of the field, and we include papers that are semantically close to that center.

Each candidate paper's SPECTER v2 embedding is compared to the centroid using cosine similarity. Papers whose similarity falls below 0.90 are excluded—they matched our keywords but are semantically distant from the core research area. This threshold was chosen empirically by reviewing papers at various similarity levels: above 0.90, papers are recognizably about AI and mental health; below, off-topic work increasingly dominates. This handles cases where keywords are ambiguous: for example, an AI paper about tropical depression tracking in weather forecasting would match a keyword search for “depression,” but its embedding would place it far from the mental health centroid, so it would be excluded.

Processing Pipeline

1. Corpus Collection

Three-stage process described above: center definition, broad keyword search, and embedding boundary filter.

2. Quality Filter

Papers must have at least 1 citation or come from an author with an h-index ≥ 10. (A researcher's h-index is the largest number h such that h of their papers have each been cited at least h times—a measure of sustained research output and impact.) This keeps the corpus focused on substantive work while avoiding penalizing recent papers from established researchers that haven't yet accumulated citations.

These thresholds are intentionally permissive—the goal is to remove clearly low-quality papers before they affect the spatial layout and topic labels, while keeping the map as comprehensive as possible. You can always raise the bar using the interactive visualization's Advanced Filters to narrow by citations, h-index, and more.

3. Enrichment

SPECTER v2 embeddings, TLDR summaries, and max author h-index are fetched from Semantic Scholar for papers that pass the quality filter.

4. Dimensionality Reduction

UMAP projects the 768-dimensional embeddings down to 2D using cosine distance, so that semantically similar papers appear near each other on the map.

5. Topic Modeling

Toponymy discovers hierarchical topic clusters from the embedding space. GPT-4o generates human-readable topic names for each cluster using paper titles and TLDR summaries (rather than full abstracts), which produces more specific and distinctive labels. Papers without a TLDR fall back to their abstract. A detail-level parameter further controls naming granularity.

6. Visualization

DataMapPlot renders the final interactive map with hover metadata, click-to-open paper links, search, and switchable colormaps.

What's Included / Excluded

Included

Matches at least one mental health keyword and one AI keyword
Meets the quality filter: has at least 1 citation, or at least one author with an h-index ≥ 10 (this avoids penalizing recent papers from established researchers)
Falls within the semantic boundary (embedding similarity above threshold)
Has a SPECTER v2 embedding
Published between 2021 and 2026

Excluded

Papers outside the semantic boundary (keyword match but topically distant)
Papers with 0 citations from authors without an established track record (h-index < 10)
Papers missing embeddings (not available from Semantic Scholar)

Semantic Scholar provides broad coverage of academic literature but is not exhaustive. Some papers—particularly very recent publications, non-English work, or papers from smaller venues—may not appear in the corpus.

Tools & Technologies

Tool	Role
Semantic Scholar API	Paper discovery, metadata, citation data, and relevance ranking
SPECTER v2	Document embeddings for semantic similarity and boundary definition
UMAP	Dimensionality reduction (768D → 2D)
Toponymy	Hierarchical topic discovery and clustering
GPT-4o	Topic name generation
DataMapPlot	Interactive visualization rendering

Using the Visualization

Reading the Map

Position: Papers near each other are semantically similar. Positions are derived from SPECTER v2 document embeddings projected to 2D via UMAP.
Node Size: Larger points indicate a higher max author h-index among the paper's authors. Sizes are log-scaled so that the long tail of very high h-index values doesn't dominate the display.
Color: Switchable via the dropdown: citations, influential citations, publication year, primary field of study, max author h-index, or centroid similarity.
Labels: Hierarchical topic names discovered automatically by Toponymy. Finer-grained labels appear as you zoom in.

Search

The search box (top-left corner) lets you find specific papers by typing any term. It matches against paper titles, authors, journal names, TLDR summaries, and abstracts, highlighting results on the map.

Filters

The Advanced Filters panel (top-left corner) lets you narrow the visible papers. Filters combine with each other—only papers that pass all active filters are shown. A live counter shows how many papers match.

Filter	What it does
Publication Year	Restrict to a specific year range. Useful for seeing how the field has evolved or focusing on the most recent work.
Citations	Filter by total citation count. Slide the minimum up to focus on widely cited work, or narrow the range to find a specific impact tier.
Influential Citations	Filter by influential citations—cases where the citing paper engages substantively with the work rather than mentioning it in passing.
Max Author H-Index	Filter by the highest h-index among a paper's authors. Slide up to focus on established research groups, or lower it to surface early-career work.
Centroid Similarity	Filter by how semantically similar a paper is to the core of "AI & mental health" research. Papers near 1.0 are squarely on-topic; papers closer to the 0.90 threshold sit at the boundary.
Field of Study	Toggle Semantic Scholar's ML-classified fields on or off. The top fields are shown individually; less common fields are grouped under “Other.”

Display Toggles

Below the filters, three toggles control map overlays. Labels (on by default) show hierarchical topic names that refine as you zoom in. Boundaries (off by default) draw colored outlines around each topic cluster, useful for seeing cluster structure when using a non-default colormap. Edge Graph (off by default) shows nearest-neighbor connections computed in the high-dimensional embedding space—edges can reveal structure that isn't always obvious from the 2D layout alone.

Field Definitions

Citations: Total number of times this paper has been cited. Learn more
Influential Citations: Citations where the citing paper uses this work substantially, rather than just referencing it in passing. Learn more
Max Author H-Index: The highest h-index among the paper's authors. A researcher's h-index is the largest number h such that h of their papers have each been cited at least h times—a measure of sustained research output and impact. This value controls point size in the visualization (log-scaled). Learn more
Primary Field of Study: An ML-classified field assigned by Semantic Scholar. The top 5 most frequent fields are shown individually; all others are collapsed to “Other.” Learn more
Topics: Hierarchical labels discovered by Toponymy. The coarsest level appears in the legend; finer-grained levels appear as you zoom in.
TLDR: An AI-generated 1–2 sentence summary of the paper from Semantic Scholar. When unavailable, a truncated abstract is shown instead. Learn more
Centroid Similarity: Cosine similarity between a paper's SPECTER v2 embedding and the centroid of the anchor paper set. Higher values indicate papers more semantically central to the AI & mental health research domain.

Citations, influential citations, and h-index use a logarithmic scale in the colormaps because their distributions are heavily right-skewed. Values are also capped at the 95th percentile so that a handful of extreme outliers don't compress the rest of the color range into a narrow band.

Contact

This project is maintained by Steven Fazzio. If you have questions, suggestions, or requests, feel free to reach out.

License

Released under the MIT License.