The study, led by Cara A. Spencer of the Computer Science Department at the University of Colorado Boulder, introduces a two-stage pipeline built on large language models and retrieval-augmented generation, or RAG. The motivation is a stubborn problem in neuroergonomics: field research in operational environments — military deployments, Antarctic stations, submarine crews, emergency response, and spaceflight analogs — is constrained by limited access, small samples, high participant burden, and costs that often exceed research budgets. Because every field study is expensive, choosing which psychological or cognitive construct to focus on is itself a high-impact scientific decision. Until now, that decision has typically been made top-down, by experts drawing on clinical taxonomies and prior literature. Spencer’s pipeline flips the logic: it lets the field data speak first, then surveys the literature to find where the evidence runs out.

Stage 1 is the construct discovery engine. The system fetched 2,521 log documents spanning 2001 to 2026, covering more than 300 discrete missions and roughly one million words of crew-authored text. Five log types were retained — Sol Summaries, Commander Reports, Journalist Reports, Mission Summaries, and Science Reports — chosen because they are most likely to contain first-person reflective content, while engineering and supply checklists were excluded. The corpus was segmented into overlapping 600-character chunks, encoded with the voyage-3-large embedding model, and persisted to a ChromaDB vector index configured for cosine similarity. Eight broad seed queries, spanning team dynamics, stress and workload, interpersonal conflict, isolation, decision-making, emotional regulation, autonomy, sleep, and physical-cognitive interfaces, swept the archive. For each sweep, the top 20 most similar chunks were retrieved, deduplicated, and independently submitted to the Claude Sonnet model with a structured extraction prompt: identify any psychological, emotional, interpersonal, or cognitive states described or implied in the passages, express each as a candidate construct with an operational definition, cite the supporting excerpts, and rate confidence.

The design choice here is technically deliberate. Because the construct vocabulary is itself the target of discovery, constructing per-construct retrieval queries would be circular. Sweeping with broad seeds and then iterating extraction across all retrieved chunks places the LLM in the role of detector rather than confirmer, so the composition of the candidate list is determined by the corpus rather than by query design. The raw labels — which included terminological redundancy such as exhaustion, fatigue, and tiredness returned as distinct labels — were frequency-filtered to retain only labels appearing in at least three independent extraction events, then canonicalized in a single call to the more powerful Claude Opus model, which grouped synonyms under canonical labels while preserving meaningful distinctions such as Crew Cohesion versus Crew Morale. Because LLM inference is non-deterministic, the canonicalization step was run 30 times, and stability selection was applied: pooled constructs were meta-clustered on cosine distance, and only clusters whose Wilson 95% lower bound on selection probability exceeded 0.50 were admitted. The result was a stable inventory of 19 canonical constructs, with a mean pairwise Jaccard similarity of 0.976 across run pairs — far above the roughly 0.80 threshold conventionally treated as indicating a stable topic-discovery configuration.

Validation was rigorous. Two trained raters, using Grounded Theory methods after achieving inter-rater reliability of κ = 0.76, independently extracted constructs from sampled logs and compared their inventory against the LLM’s output and against NASA’s Human Research Program Human Factors and Behavioral Performance Evidence reports — the current gold standard for spaceflight behavioral risk. The correspondence was strong: 57.9% of the 19 constructs received full convergent agreement from both reference standards, and all four highest-frequency constructs — Crew Cohesion (26 passages), Emotion Management/Regulation (17), Sleep Quality/Disruption (14), and Cognitive Fatigue/Performance Degradation (11) — were rated as fully corresponding by both. Human raters also confirmed that the RAG system was comprehensive (all 2,521 logs fetched), relevant (90% of evidence in a sampled benchmark), and faithful (94% of identified constructs answered the prompt). Notably, a few constructs surfaced by the pipeline — acute stress response, self-awareness, competitive orientation, and other-oriented emotional concern — were less represented in the NASA taxonomy, suggesting the bottom-up approach can catch phenomena that top-down frameworks underweight.

Stage 2 then turned the construct list into an evidence gap matrix. For each of the 19 constructs, structured queries — augmented with a generated synonym list — were submitted to three bibliographic databases: PubMed via the E-utilities API, Semantic Scholar via its bulk search endpoint, and IEEE Xplore via its Metadata API. Retrieval was configured to be exhaustive rather than capped, using source-specific pagination to page the complete match sets. Deterministic post-retrieval filters removed non-human species studies, purely clinical populations, and records under age 18, and a relevance filter required a construct token in the title or abstract. The pipeline fetched 49,172 abstracts in just 17 minutes; after deduplication and filtering, 28,541 abstracts were evaluated by the LLM. Each abstract was scored along four evidence dimensions that together characterize the translational pipeline from laboratory neuroscience to deployable field monitoring: whether the construct has been studied under ecologically valid field conditions, whether a published neural correlate exists, whether a validated biosensor protocol exists outside the laboratory, and whether the construct has been used as a control signal in a working brain-computer interface.

Hallucination mitigation was built in at every step. Every reference cited by the LLM was automatically cross-referenced against the per-construct retrieval database using a two-stage title-matching procedure, flagging 3.6% of citations as unverified — a very low rate of fabricated bibliographic entries. Human reviewers examined a dimension-stratified sample of 200 abstracts blind to the LLM’s labels; human-human agreement on the evidence dimensions ranged from κ = 0.82 to κ = 0.95, and human-LLM agreement, initially only a pooled κ = 0.56, was iteratively improved to κ = 0.74 through prompt revision. An expert reviewer also re-evaluated all matrix cells containing zero by manually searching the browser versions of the databases — and the replicated searches produced the same results, confirming that the near-empty cells were real rather than artifacts of the pipeline.

The headline finding is structural, and it is striking. Crew-authored text most heavily emphasized social and affective phenomena — crew cohesion anchored the most prominent hub of the co-occurrence graph, connected across multiple domains. Yet the evidence gap matrix shows these are precisely the phenomena the existing literature is least equipped to monitor. The deficit follows two compounding gradients. The first is construct domain: evidence attenuates sharply from physical and cognitive constructs like fatigue and sleep disruption, which carry thousands of supporting abstracts across every dimension, to social and motivational constructs like Crew Morale, Agentic Leadership Motivation, and Privacy Deprivation, which return single-digit or zero counts. The second is translational stage: within a construct, counts fall steeply from establishing that a neural correlate exists toward demonstrating a deployable biosensor or BCI protocol. Emotion Management/Regulation, for example, is supported by 3,122 neural-correlate abstracts but only 133 for a validated biosensor protocol and 167 for BCI implementation; Self-Awareness collapses from 458 neural-correlate abstracts to 14 and 5, respectively. The bottleneck, in other words, sits at the translation from laboratory correlate to field-deployable protocol — not at the absence of a neuroscientific basis. The two gradients intersect at the social and affective constructs most central to crew accounts, driving their deployment readiness toward zero. The authors note a plausible physical explanation: wearable fNIRS penetrates only about 3 centimeters and does not reach the amygdala, so many affective processes abundant in laboratory neuroimaging evidence simply cannot carry through to wearable protocols.

The comparison with human effort is dramatic. A two-person literature team, searching each database manually for two hours, evaluated just 31 abstracts; the pipeline evaluated tens of thousands in minutes while remaining traceable to specific retrieved passages. The false-negative analysis indicated 99% of filtered-out abstracts were correctly removed. Yet the study is equally clear about the limits: the pipeline’s outputs depend on the choice of LLM, embedding model, and prompting regime; retrieval was abstract-only and restricted to three databases; the corpus sweep did not enforce balance across missions or contributors, in principle allowing prolific log writers to over-weight the candidate pool; and the evidence cutoff date will become progressively stale. The final construct inventory was still reviewed and edited by a human, and the authors explicitly position the system as a powerful but fallible first-pass tool whose outputs are credible enough to inform research priority decisions only when paired with expert scrutiny.

The implications reach far beyond Mars analogs. The pipeline is corpus-agnostic at Stage 1 and construct-list-agnostic at Stage 2: any operational environment that produces structured written records — military after-action reviews, emergency response logs, submarine deployment journals, even financial trading-desk diaries — can in principle serve as input. The evidence gap matrix also distinguishes two qualitatively different research opportunities: constructs with strong neural correlates but sparse biosensor and BCI evidence mark opportunities for protocol translation, where a demonstrated correlate awaits engineering into a deployable measurement, while the near-empty social constructs mark where foundational neurophysiological work is still required before translation can begin. Fatigue, by contrast, emerges as the lowest-risk target — mature across every dimension yet under-studied specifically in ICE settings. For a field preparing for long-duration spaceflight to the Moon and Mars, where detecting emerging behavioral health risks without relying solely on self-report could be mission-critical, the study offers a concrete, quantified agenda. It also offers a template for how generative AI, deployed with retrieval grounding, stability analysis, citation verification, and human-in-the-loop checkpoints, can compress months of manual synthesis into weeks — while keeping scientists firmly in command of the conclusions.

,

Subject of Research: A two-stage RAG-LLM pipeline for roadmapping neuroscience research in operational environments

Article Title: A two-stage RAG-LLM pipeline for roadmapping neuroscience research in operational environments

Article References: A two-stage RAG-LLM pipeline for roadmapping neuroscience research in operational environments. (n.d.). https://doi.org/10.3389/fnhum.2026.1884892

Image Credits: AI Generated

DOI: 10.3389/fnhum.2026.1884892

Keywords: two-stage, RAG-LLM, pipeline, roadmapping, neuroscience, research, operational, environments, scientific research