Introduction and methods

1. Introduction

From Proust’s literary observation to the nineteenth-century concept of the “smell brain” (rhinencephalon), the link between odour, emotion and memory was widely accepted long before tools existed to measure it. The past three decades have changed the foundations of the question in three ways. First, molecular biology identified the odorant receptor gene family [40] and, more recently, the first atomic structure of a human odorant receptor [42], showing that the periphery already performs non-linear computation before the signal reaches the brain [8]. Second, multivoxel pattern decoding (MVPA), intracranial EEG (iEEG) and single-cell recording in humans make it possible to ask directly which regions encode identity and which encode valence [11–13, 49, 51]. Third, the boom in the essential-oil market and in scent marketing has produced hundreds of applied trials, most with small samples and difficult to blind [101, 104].

As a result, current knowledge is fragmented: mechanistic reviews rarely assess the certainty of applied evidence, while applied meta-analyses rarely place their results within a mechanistic framework. This review aims to connect the two, with three questions:

1. Q1 — Mechanism: How is the odour signal transformed from the receptor to the amygdala, OFC and entorhinal–hippocampus, and what information does each region encode?
2. Q2 — Magnitude: How large are the characteristic behavioural effects (odour-evoked autobiographical memory, reactivation during sleep, emotional chemosignals) when measured quantitatively?
3. Q3 — Application: At what level of certainty are common applied claims (aromatherapy, olfactory training, scent marketing) supported?

The scope is limited to the human main olfactory system, using animal data where needed for causal mechanisms; the accessory (vomeronasal) system and taste are outside the scope.

2. Methods

2.1 Design

A narrative systematic review with descriptive quantitative synthesis, reported according to the SWiM (Synthesis Without Meta-analysis) guideline and following the PRISMA 2020 structure where applicable. The review did not pre-register a protocol and the set of studies is not the result of a comprehensive database search; this is the main limitation compared with a standard systematic review and is discussed in section 4.5.

2.2 Sources and selection

Set 1 comprises 39 items from an existing reference list, re-verified against DOI/PubMed (17 items needed corrections to authors, journal or title). Set 2 comprises 80 papers from a targeted search across four domains: (A) receptor and olfactory cortex foundations; (B) amygdala, valence, chemosignals and breathing rhythm; (C) hippocampus, memory, sleep, development and ageing; (D) clinical and practical applications. Priority was given to open-access full texts, highly cited foundational papers, human studies, systematic reviews and meta-analyses. Every DOI was verified via Crossref/OpenAlex; open-access status via Unpaywall and the PMC ID converter. The flow diagram is in Figure 1.

2.3 Full-text retrieval

Full texts were obtained in the following order: the publisher’s open version, PMC (including the OAI-PMH endpoint), university repositories, author manuscripts, subscription databases (ClinicalKey), and manual browser access for sites that block automated access. No bot-verification mechanism was circumvented. For papers with only an abstract, the notes add secondary descriptions from open reviews and mark the source of each figure.

2.4 Data extraction

Extraction was done in two tiers. Tier 1: each paper was read in full and summarised in an 11-item note (question, context, methods, results with figures, mechanism, limitations, conclusion, figures–tables, terminology, background literature, implications), totalling ~311,000 words. Tier 2: from the notes, data were coded into a 25-field schema (design, species, sample size, intervention, comparator, outcome, main finding, effect size with CI/k/N/I², mechanism, limitations, risk of bias, level of evidence, funding, implications). Result: 277 effect-size rows from 75 papers. Extraction was carried out by a language-model-assisted analysis workflow, split into independent batches; there was no independent double extraction by two people.

2.5 Quality assessment

The level of evidence of each paper was graded on the OCEBM scale (1a: systematic review/meta-analysis of RCTs or controlled experiments — some 1a papers in this set pool non-randomised studies [111, 113] or marketing experiments [117], so level 1a here is broader than the original definition; 1b: RCT; 2: cohort; 3: cross-sectional/case-control; 4: small experiment; 5: mechanistic, animal, expert opinion). Risk of bias was recorded from the original paper’s assessment (Cochrane RoB, Jadad) where available; otherwise, a summary assessment was made based on sample size, blinding, funding and signs of publication bias. Certainty by outcome (Table 6) followed GRADE principles: starting high for RCTs/randomised experiments and low for observational studies or non-randomised interventions; downgrading for risk of bias, inconsistency, indirectness, imprecision and publication bias; upgrading for large effects or a dose–response relationship. The GRADE assessments are those of the authoring team, not of the original papers’ authors.

2.6 Synthesis

Studies were grouped by processing stage (periphery → olfactory bulb/piriform → amygdala → OFC → entorhinal–hippocampus) and by application domain. No further meta-analysis was pooled because outcomes, scales and designs were heterogeneous; instead, 19 pooled and single estimates were placed on the same standardised scale (g/d/SMD) to compare magnitude and precision (Figure 8). Correlation coefficients were converted using d = 2r/√(1−r²). Vote counting by statistical significance was not used.

2.7 Data quality control

Every inconsistency detected within and between papers was recorded in a log (Appendix C), including suspected minus-sign errors, numerical inconsistencies between the main text and tables, and incorrect secondary citations. Where a point estimate conflicted with its confidence interval, the value consistent with the CI was used and flagged.

Scientific reference material, not medical advice. Figures are redrawn from published data, not copied from the original figures. GRADE and risk-of-bias ratings are the compilers’ own. Every number traces back to detailed reading notes on each paper.