Evidence Sources & Methodology

ElfieCare Evidence AI is built on a curated, governed medical knowledge system — not a general-purpose search engine. Every response is grounded in verifiable external sources, classified by evidence strength, and kept current through structured review.

This page explains what powers it.

‍

Our Principles

1. External truth only. All knowledge originates from verifiable external sources. The system does not generate medical facts.
2. Governance before usage. Nothing is surfaced to physicians until it has been structured, classified, and reviewed.
3. Honesty over coverage. When evidence is absent, weak, or conflicting, the system is designed to say so — clearly and without speculation.
4. Traceability. The system is designed to log the evidence basis, confidence level, and resolution path for each response, supporting audit and clinical governance.

‍

1. What We Index

Evidence AI draws from a curated registry of medical sources, organized into three layers.

1.1 Global Evidence Backbone

These universal sources underpin every clinical query, regardless of a physician's location.

1.1.1 Open-Access Literature and Trial Registries

1. PubMed / MEDLINE — the world's largest index of biomedical literature, providing structured abstracts across all indexed journals.
2. PubMed Central (PMC) — millions of full-text, open-access research articles.
3. ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform (ICTRP) — active and completed trial data for clinical currency.
4. Cochrane Library — structured abstracts and plain-language summaries from the gold standard in systematic reviews.

1.1.2 Global Health Authorities

1. World Health Organization (WHO) — guidelines, disease burden data, vaccination schedules.
2. U.S. Centers for Disease Control and Prevention (CDC) — public health guidance and immunization schedules.
3. National Institutes of Health (NIH) — research summaries and clinical guidance.
4. U.S. Food and Drug Administration (FDA) — drug labels, safety communications, and black box warnings.
5. European Medicines Agency (EMA) — European drug safety and approval information.

1.1.3 Open-Access Peer-Reviewed Journals

JAMA Network Open, Journal of Medical Internet Research (JMIR), BMC Medicine, PLOS Medicine, Frontiers in Medicine, Diabetes Care, and others that publish under open licenses without commercial restrictions.

For leading journals where full-text access requires licensing, we index structured abstracts available through PubMed. These contain key findings, study design, effect sizes, and clinical conclusions. Each abstract-sourced item is transparently marked so that the system can disclose the limitation.

1.2 International Specialty Guidelines

Evidence AI indexes clinical practice guidelines from major international medical associations and specialty bodies where their published guidelines are openly available.

1. American Diabetes Association (ADA) — Standards of Medical Care, published annually in open access.
2. Infectious Diseases Society of America (IDSA) — clinical practice guidelines.
3. KDIGO (nephrology), GINA (asthma), GOLD (COPD), WHO mhGAP (mental health), World Gastroenterology Organisation (WGO) — world-standard specialty guidelines published openly.
4. International Federation of Gynecology and Obstetrics (FIGO), American College of Physicians (ACP), World Federation of Societies of Anaesthesiologists (WFSA).

Some specialty society guidelines require commercial licensing agreements. Where these are not yet in place, the system does not ingest the content. Where joint publications or consensus reports are available through open-access channels, those specific items are included.

1.3 Local Regulatory and Guideline Sources

Evidence AI is geo-aware. It uses the physician's country and specialty to prioritize local regulatory and guideline sources over global ones.

We index sources from national health authorities, drug regulators, and local clinical guideline bodies across our launch markets, including:

1. United Kingdom — NICE guidelines, NHS clinical pathways, MHRA drug safety data.
2. France — Haute Autorité de Santé (HAS), ANSM, Ministère de la Santé.
3. Australia — Department of Health and Aged Care, TGA, PBS, Australian Diabetes Society.
4. United States — CDC, FDA, NIH, USPSTF.
5. Brazil — Ministério da Saúde, ANVISA, CONITEC, Sociedade Brasileira de Cardiologia (SBC).
6. Spain — AEMPS, guiasalud.es, SEOM.

Additional local sources cover markets across Latin America, the Middle East and North Africa, Sub-Saharan Africa, and Southeast Asia. The full local source list is maintained internally and expanded as new markets are supported.

‍

2. What We Don't Index

We are transparent about our boundaries.

2.1 Licensed Content

A number of important medical guidelines and reference databases are protected by copyright or licensing terms that restrict commercial use. We respect these restrictions. Where licensing agreements are not in place, the content is not ingested — regardless of its clinical value.

We are actively pursuing partnerships to expand coverage where licensing permits.

2.2 Content That Does Not Meet Quality Standards

The system does not index:

1. Non-peer-reviewed content such as blog posts, social media commentary, or opinion pieces without evidence grading.
2. Preprints that have not undergone peer review.
3. Marketing materials, regardless of origin.
4. Internal hospital protocols that have not been explicitly shared and qualified through our workspace-level guideline feature.

Every source entering the system must pass a Source Qualification process that verifies legitimacy, authority class, document type, and publication freshness before extraction begins.

‍

3. How We Build the Hierarchy of Evidence

Not all evidence carries equal weight. A meta-analysis of multiple randomized controlled trials is stronger than a single case report. A country-specific guideline takes precedence over a global recommendation for a doctor in that country. Evidence AI encodes these distinctions structurally.

3.1 Evidence Tiers

Every knowledge object is classified into one of five evidence tiers:

Tier | Type | Examples

‍T1 | Clinical guidelines — region-matched or global | National authority guidelines, WHO recommendations

‍T2 | Systematic reviews and meta-analyses | Cochrane reviews, high-quality evidence syntheses

‍T3 | Primary trials | Randomized controlled trials

‍T4 | Observational evidence | Cohort studies, case-control studies

‍T5 | Consensus and lower-tier support | Expert opinion, case series, consensus statements

3.2 GRADE Framework

Where applicable, evidence objects are further characterized using the GRADE (Grading of Recommendations, Assessment, Development, and Evaluations) framework, which captures evidence certainty, strength of recommendation, and individual quality dimensions such as risk of bias, inconsistency, indirectness, and imprecision.

3.3 Resolution Rules

When a query returns evidence from multiple sources, the system applies structured resolution rules:

1. Local over global — a local guideline is prioritized over a global one for a physician in that country.
2. Newer over older — when sources are at the same evidence level, more recent guidance is preferred.
3. Higher evidence level over lower — by tier and level of evidence.
4. Disagreement is preserved — when conflict between strong sources cannot be resolved, both viewpoints are presented with full citations. The system is designed to surface disagreement rather than silently flatten it.

3.4 Confidence Scoring

Each knowledge object carries a composite confidence score derived from two components:

1. Quality — evidence tier, GRADE characterization, governance completeness, and whether conflicts exist.
2. Freshness — source age relative to its review cadence, review recency, lifecycle state, and whether the source organization has published an update.

This score determines retrieval behavior:

• High confidence — cited with confidence. Language: "Evidence supports…"
• Moderate confidence — cited with explicit caveats. Language: "Evidence suggests…"
• Low confidence — surfaced only when stronger evidence is unavailable. Language: "Limited evidence indicates…"
• Insufficient confidence — withheld from clinical outputs and queued for review.

When evidence is insufficient or absent, the system is designed to state that explicitly and offer structured query refinements — rather than generate a speculative answer.

‍

4. How We Keep It Current

Medical evidence has a shelf life. Guidelines are updated, new safety alerts are issued, and once-standard treatments are superseded. A static knowledge base decays silently. Evidence AI addresses this through five interlocking mechanisms.

4.1 Review Cadence

Every knowledge object is assigned a review cadence based on clinical risk and expected volatility:

Class | Frequency | Scope

‍A | 3–6 months | High-risk, fast-changing evidence — therapy guidelines, drug safety alerts

‍B | 6–12 months | Core clinical guidelines and common conditions

‍C | 12–24 months | Stable foundational knowledge

‍D | On-demand | Low-priority or historical content

4.2 Event-Triggered Reviews

Scheduled cadence is supplemented by event-driven review triggers:

1. A source organization publishes a new guideline version.
2. A drug regulatory authority issues a safety update.
3. Conflicting high-level evidence emerges.
4. The system detects a contradiction between existing knowledge objects.
5. A source retraction or published erratum is identified — triggering an immediate lifecycle cascade.

4.3 Freshness Decay

The freshness component of each object's confidence score decays over time, calibrated to its review cadence. Fast-moving evidence (Class A) decays more rapidly than stable knowledge (Class C/D). If an object's confidence falls below a safe threshold, the system is designed to automatically flag it for review and restrict its use in clinical outputs.

4.4 Lifecycle States

Each knowledge object exists in one of four states:

1. Active — current, valid, eligible for retrieval.
2. Caution — usable but flagged due to age, partial conflict, or pending review.
3. Deprecated — replaced by stronger or newer evidence, or source retracted.
4. Retired — removed from active use, retained for audit.

4.5 Clinical Review and Contradiction Detection

Evidence objects are maintained by qualified clinical reviewers. High-risk objects are subject to dual review. Disagreements between reviewers are escalated and must be resolved before an object is approved for use. Review events are recorded for auditability.

The system also monitors for conflicts between knowledge objects. When a new guideline contradicts an existing one, the contradiction is flagged and routed for review. At the physician-facing layer, unresolved conflicts are presented with both viewpoints and citations — the system is designed to let the clinician decide.

‍

For professional use only. ElfieCare Evidence AI synthesizes medical literature based on clinical parameters. It does not provide medical diagnoses or treatment mandates. Final clinical judgment remains with the healthcare provider.

‍