---
title: 'AgentSLR: Automated SLR in Epidemiology'
url: https://www.emergentmind.com/topics/agentslr
type: topic
---

# AgentSLR: Automated SLR in Epidemiology

Searching arXiv for the cited AgentSLR paper and closely related SLR automation work to ground the article.
AgentSLR is an open-source agentic pipeline for automating systematic literature reviews in infectious disease epidemiology, spanning article retrieval, title and abstract screening, PDF-to-Markdown conversion, full-text screening, structured data extraction, and evidence-grounded report synthesis. Applied to epidemiological reviews of nine WHO-designated priority pathogens and validated against expert-curated ground truth, it achieves performance comparable to human researchers while reducing review time from approximately 7 weeks to 20 hours, a 58x speed-up [2603.22327].

## 1. Scope, domain, and review target

AgentSLR is designed for systematic literature reviews in epidemiology, where the review object is not merely bibliographic relevance but extractable quantitative evidence about transmission dynamics, temporal parameters, severity, seroprevalence, risk factors, outbreak characteristics, and transmission-model structure. The pipeline was applied to nine WHO priority pathogens: Marburg virus, Ebola virus, Lassa fever virus, SARS-CoV-1, Zika virus, MERS-CoV, Nipah virus, Rift Valley fever virus, and CCHF virus [2603.22327].

The system is anchored to the Pathogen Epidemiology Review Group (PERG) review framework. PERG provides the expert-curated ground truth used for screening and extraction evaluation, with REDCap-based schemas for parameters, models, and outbreaks, and downstream distribution via R packages. Evaluation is restricted by open-access availability: only approximately 26.2% of PERG-considered articles are available to AgentSLR after deduplication and open-access filtering, so screening evaluation is reported on seven pathogens and extraction evaluation on four pathogens, with outbreak extraction evaluated only on Lassa and Zika [2603.22327].

The target review objects are epidemiologically specific. Parameter extraction covers classes such as attack rate, doubling time, growth rate, mutations, human delays, overdispersion, relative contribution, reproduction number, severity, seroprevalence, and risk factors. Model extraction captures structural properties such as model type, compartmental type, stochastic versus deterministic formulation, spillover, assumptions, interventions, and reproducibility metadata. Outbreak extraction captures dates, geography, counts, detection mode, epidemiological context, and demographic disaggregation [2603.22327].

## 2. Pipeline architecture and workflow

AgentSLR is organized as a six-stage workflow: article search and retrieval; title and abstract screening; PDF-to-Markdown conversion; full-text screening; data extraction; and report generation with self-refinement. The agentic character of the system lies chiefly in the reasoning LLM that orchestrates multi-step operations within screening, extraction, and synthesis, with later stages conditioned on structured outputs and manifests produced earlier [2603.22327].

Article search and retrieval use OpenAlex, PubMed, and Europe PMC. Search is based on a central Boolean template covering domains such as transmissi*, epidemiolog*, model*, severity, delays, heterogeneity, reproduction numbers, serology, evolution, outbreaks, and risk factors, augmented by pathogen-specific identifiers. Deduplication uses a hierarchical five-level procedure over DOI, PMID, PMCID, OpenAlex ID, and normalized title plus year. PDF retrieval then queries multiple open-access sources, including OpenAlex direct PDF, Europe PMC full text, Unpaywall, and DOI-based OpenAlex fallback, with HTTP streaming, PDF validation, and parallel downloads with caching and checkpointing [2603.22327].

Title and abstract screening uses a reasoning LRM with ScreenPrompt-style prompting. The model receives study objectives, explicit inclusion and exclusion criteria, article title and abstract, and is instructed to think step by step, returning `<decision>INCLUDE</decision>` or `<decision>EXCLUDE</decision>`. Full-text screening preserves the same logic but operates on Markdown-converted full text and imposes a stricter requirement that the paper contain extractable quantitative parameters or models. The Markdown conversion stage itself is tool-based rather than LLM-driven: PDFs are split into pages, rendered to high-resolution images, and converted by `mistral-ocr-2512` into Markdown with section hierarchy, equations as LaTeX, and tables as HTML or structured Markdown [2603.22327].

Report generation is evidence-grounded rather than free-form. Programmatic routines first compute dataset statistics, tables, and figures, and construct a Markdown draft plus manifest. The LLM then produces a living review conditioned on this evidence packet and explicit rubrics for data fidelity, outbreak focus, figure and table presence, traceability, clarity, completeness, and proper use of `> AI-Interpretation:` blocks. Self-refinement proceeds for up to \(K = 5\) critique-and-revise iterations, with hard constraints that required figures appear at least once and all tables preserve unchanged values [2603.22327].

## 3. Agentic extraction procedures and schema design

The most technically distinctive part of AgentSLR is the extraction layer, which decomposes parameter, model, and outbreak extraction into staged tool-calling pipelines with strict JSON schemas and validation. Across all three families, the recurring pattern is article-level flagging, targeted extraction, validation, and provenance linking [2603.22327].

Parameter extraction begins with parameter-class screening. For each class, the model decides whether the article estimates that parameter from actual data and, if so, calls a tool returning quoted summary snippets. Each summary is then processed by a class-specific extraction prompt that fills typed fields such as values, units, numerators and denominators, uncertainty bounds, method, and value type. A separate call extracts population context, including sex, population group, sample type, sample size, age bounds, countries, locations, and moment in outbreak. If a paper reports many disaggregated estimates, the PERG “rule of three” triggers aggregation into ranges. A further pass links extracted items back to textual provenance [2603.22327].

Model extraction follows a similar logic. First the system flags whether the article contains dynamic transmission models, explicitly excluding pure statistics, regression-only forecasting, and risk-factor studies without transmission modeling. It then extracts one structured record per distinct model. The schema includes `model_type`, `compartmental_type`, `stoch_deter`, `theoretical_model`, transmission-route fields, `spatial_model`, `spillover_included`, assumptions, interventions, and reproducibility fields such as `code_available`, `coding_language`, `is_data_used_available`, and `readme_included`. Validation rejects inconsistent calls, such as a non-compartmental model paired with a compartmental subtype other than “Not compartmental” [2603.22327].

Outbreak extraction targets concluded real-world outbreak events with bounded time periods and case counts. The schema includes start and end day, month, and year; `outbreak_duration_months`; `outbreak_is_currently_ongoing`; country and fine-grained location; source; mode of detection; pre-outbreak status; multiple case-count fields; `asymptomatic_transmission_described`; population size; and sex-disaggregated counts. The extractor is constrained by pathogen-specific rules, such as minimum-case thresholds for some pathogens and single-case allowances for others [2603.22327].

The evaluation formalism for extraction is explicitly structured. For field-level matching, AgentSLR constructs a bipartite matching between PERG extractions and AgentSLR extractions using a similarity function
\[
s(E,\hat{E}) = \sum_{k \in \mathcal{F}} w_k \cdot d_k(E[k], \hat{E}[k]),
\]
where \(d_k\) is Jaccard similarity for multi-valued fields or exact match for single-valued fields, and matching is solved with Jonker-Vollgenant via `linear_sum_assignment`. Precision, recall, and \(F_1\) are then computed from aggregated field-level \(TP\), \(FP\), and \(FN\) counts [2603.22327].

## 4. Empirical evaluation and task performance

For screening, AgentSLR reports standard binary-classification metrics,
\[
\text{Precision} = \frac{TP}{TP + FP}, \quad
\text{Recall} = \frac{TP}{TP + FN}, \quad
F_1 = \frac{2\,\text{Precision}\,\text{Recall}}{\text{Precision}+\text{Recall}},
\]
with macro-averaging used because of label imbalance. With `gpt-oss-120b`, title and abstract screening achieves overall precision \(0.79\), recall \(0.72\), and \(F_1 = 0.74\), with pathogen-level \(F_1\) ranging from \(0.69\) for Marburg to \(0.78\) for MERS [2603.22327].

Full-text screening is evaluated in three operating modes. In the fully automated two-stage setting, where AI abstract screening gates AI full-text screening, the system obtains precision \(0.75\), recall \(0.81\), and \(F_1 = 0.77\). When conditioned on human abstract decisions, full-text screening improves to precision \(0.83\), recall \(0.92\), and \(F_1 = 0.87\). Direct AI full-text screening without abstract gating reaches recall \(0.89\) but lower precision \(0.68\), yielding \(F_1 = 0.73\) and requiring approximately \(2.3\times\) more OCR and runtime [2603.22327].

Extraction performance is heterogeneous across task families. For parameter extraction, averaged across Lassa, Ebola, SARS, and Zika, flagging reaches \(P = 0.51 \pm 0.07\), \(R = 0.92 \pm 0.06\), \(F_1 = 0.66 \pm 0.06\); counts reach \(P = 0.83 \pm 0.10\), \(R = 0.47 \pm 0.09\), \(F_1 = 0.59 \pm 0.07\); and field-level extraction reaches \(P = 0.52 \pm 0.03\), \(R = 0.57 \pm 0.04\), \(F_1 = 0.54 \pm 0.02\). Method fields and single-type uncertainty fields are strong, whereas value fields and population-context fields are materially weaker [2603.22327].

Model extraction is stronger. Averaged across the same pathogens, flagging reaches \(P = 0.90 \pm 0.04\), \(R = 0.91 \pm 0.05\), \(F_1 = 0.91 \pm 0.04\); counts reach \(P = 0.52 \pm 0.05\), \(R = 0.99 \pm 0.01\), \(F_1 = 0.68 \pm 0.04\); and field-level extraction reaches \(P = 0.63 \pm 0.04\), \(R = 0.74 \pm 0.02\), \(F_1 = 0.67 \pm 0.03\). Core structural fields such as `model_type`, `stoch_deter`, `theoretical_model`, and `code_available` are comparatively strong, whereas assumptions and interventions are harder [2603.22327].

Outbreak extraction, evaluated on Lassa and Zika, shows the highest field-level extraction performance: flagging reaches \(P = 0.63 \pm 0.06\), \(R = 0.76 \pm 0.05\), \(F_1 = 0.61 \pm 0.09\); counts reach \(P = 0.66 \pm 0.17\), \(R = 0.72 \pm 0.28\), \(F_1 = 0.69 \pm 0.22\); and field-level extraction reaches \(P = 0.85 \pm 0.00\), \(R = 0.76 \pm 0.02\), \(F_1 = 0.79 \pm 0.01\). Country fields are perfect, temporal fields are strong, and fine-grained location fields are weaker [2603.22327].

## 5. Model comparisons, time compression, and cost structure

AgentSLR was evaluated with `gpt-oss-120b`, GPT-5.2, Kimi-K2.5, GLM-4.7, and DeepSeek-V3.2 for the reasoning and tool-calling stages; Anthropic Claude Opus 4.5 and Sonnet 4.5 could not be used because streaming refusals were triggered by safety filters. The central finding is that performance on systematic review tasks is driven less by model size or inference cost than by distinctive model capabilities, especially context handling and robust tool use [2603.22327].

On title and abstract screening, Kimi-K2.5 achieves macro-\(F_1 = 0.77\), followed by `gpt-oss-120b` at \(0.74\), GLM-4.7 at \(0.72\), GPT-5.2 at \(0.65\), and DeepSeek-V3.2 at \(0.62\). On full-text screening, `gpt-oss-120b` is best at \(0.77\), followed by Kimi-K2.5 at \(0.71\), GLM-4.7 at \(0.69\), GPT-5.2 at \(0.59\), and DeepSeek-V3.2 at \(0.52\). For extraction, the best overall model varies by task: Kimi-K2.5 is best for parameter extraction at average \(F_1 \approx 0.63\), GLM-4.7 is best for model extraction at \(F_1 \approx 0.85\), and GPT-5.2 is slightly best for outbreak extraction at \(F_1 \approx 0.77\) [2603.22327].

The time compression is substantial. For a typical review averaged across Ebola, Lassa, SARS, and Zika, human effort is estimated at approximately \(385.1\) hours: \(114.2\) hours for title and abstract screening, \(73.5\) hours for full-text screening, and \(197.5\) hours for extraction. AgentSLR completes the same workflow in approximately \(20.0\) hours, including article retrieval, title and abstract screening, OCR, full-text screening, and data extraction. This corresponds to a \(19.3\times\) labour-time reduction and, under realistic human working hours, a \(58\times\) reduction in calendar days [2603.22327].

Cost is dominated by parameter extraction. Per article, `gpt-oss-120b` uses roughly \(2.3\)K input and \(1.2\)K output tokens for title and abstract screening, but approximately \(510\)K input and \(19.8\)K output tokens for parameter extraction. At the review level, `gpt-oss-120b` costs about \(\$13.9\) per pathogen SLR via OpenRouter and about \(\$50\) when OCR and overhead are included, whereas GPT-5.2 is reported at about \(\$1{,}348\). `gpt-oss-120b` thus occupies the lowest-cost point among the compared models while remaining competitive in accuracy [2603.22327].

## 6. Human oversight, limitations, and position within SLR automation

AgentSLR is explicitly presented as an assistant rather than an autonomous replacement for expert judgment. A beta annotation tool supports dual-pane human review, with the source article on one side and structured extractions plus highlighted evidence on the other, allowing experts to accept, revise, or reject each item. Expert-rated field-level accuracy exceeds strict exact-match scores, reaching average values of \(0.77\) for parameters, \(0.83\) for models, and \(0.80\) for outbreaks, with competence scores on a 1–7 scale of \(4.2\), \(2.8\), and \(3.9\), respectively [2603.22327].

The main limitations are concrete. Coverage is restricted to open-access, predominantly English-language literature. Exact-match evaluation can understate practical usefulness, especially when near-equivalent values or partially correct extractions are clinically or scientifically serviceable. The orchestration is staged and tool-centric rather than a fully autonomous meta-agent that revises search strategies or defines new extraction tools, and the system does not itself perform statistical meta-analysis; it produces normalized datasets, descriptive statistics, figures, and narrative synthesis ready for downstream analysis [2603.22327].

Within the broader SLR-automation landscape, AgentSLR occupies the end-to-end evidence-synthesis position. A distinct line of work focuses on evaluating the quality of systematic reviews with PRISMA-aligned multi-agent systems rather than executing retrieval, screening, and extraction [2509.17240]. Another line, SWARM-SLR AIssistant, embeds the first SWARM-SLR stage in a modular, agent-based interface with persistent storage and a tool registry, but remains primarily a workflow assistant rather than a complete extraction-and-reporting pipeline [2603.05177]. This suggests a useful conceptual distinction between agentic systems for review-methodology support and agentic systems for full evidence-synthesis execution.

Source: https://www.emergentmind.com/topics/agentslr