Check whether claims match experiments, exact reported results, and stated limitations.Know what the science actually supports before you trust the answer.
Press Enter β΅ to review paper
Explore by Goal
"We live in a society exquisitely dependent on science and technology, in which hardly anyone knows anything about science and technology."
- Carl Sagan
Quick Answer
Copied
Core finding (with skepticism)
The paper uses ensemble boosted regression trees (BRT) to infer which spatial scales of deforestation-related landscape metrics best predict household-level Plasmodium knowlesi occurrence, finding a strongest scale peak for βcleared land within 1 km,β and scale-dependent effects for forest loss fragmentation at larger radii.
Long Answer
Predictive analysis across spatial scales links zoonotic malaria to deforestation
Proceedings of the Royal Society B (DOI: 10.1098/rspb.2018.2351)
What the study actually does (mechanism-light, scale-heavy)
Outcome: household-level P. knowlesi occurrence using PCR-confirmed cases as presence points.
Absences: absence households are sampled from a cross-sectional survey sampling frame, with multiple spatial and plausibility filters to avoid obvious sampling bias.
Predictors: 16 satellite/terrain/accessibility variables summarized across 11 spatial scales around each household (using maximum overland distance buffers rather than circular ones).
Model class: ensemble boosted regression trees (BRT) with 10-fold cross-validation; variable importance across scales is interpreted via relative variable importance (RVI).
Important: this is prediction and scale inference, not direct mechanistic proof (the authors explicitly narrow plausible mechanisms rather than demonstrate them).
Visuals from the paper (numbers we do know)
Model discrimination (AUC)
Presence/absence data counts used
Counts: 206 presence points and 1324 absence points are reported; inclusion/exclusion filters and island subsetting are also described.
What the scale peaks imply (and where skepticism is warranted)
Main reported scale-dependent signals
Cleared land within 1 km: the paper reports the proportion of cleared land within 1 km has the strongest and best-defined peak of predictive influence on P. knowlesi occurrence.
Large-scale fragmentation: fragmentation of deforested/forest-loss patterns influences P. knowlesi probability most strongly at larger spatial scales (~4β5 km).
Small-scale topography & peridomestic change: aspect (1 and 2 km), slope (~0.5 km), and canopy regrowth (0.5 km) are reported as important at small scales.
Mechanistic story offered (still inferential): the authors propose that movement/density processes of humans, macaques, and vectors across distances could make certain landscape geometries predictive at corresponding radii; however, the paper emphasizes mechanism uncertainty and uses scale as a constraint rather than proof.
Critical appraisal: what could bias βscale peaksβ?
The authors themselves caution that some scale peaks could reflect variance availability and metric artifacts, especially where RVI tracks variance across correlated scales and where square grid geometry may overestimate perimeter length at small scales.
Absence points are not simply random non-cases; they are drawn from survey villages with constraints to avoid proximity bias and urban locations, but the paper still relies on the assumption that these absences appropriately represent the underlying risk landscape.
3) Correlation vs causation at multiple scales
BRT importance profiles identify correlational scale relationships between landscape summaries and case occurrence; the authors argue this supports plausible mechanism classes, but causal mediation (vectors vs macaques vs human behavior) is not directly tested in this paper.
4) Feature confounding (roads/clinics vs environment)
Adding non-scalable predictors (distance to clinic and road) only slightly improves AUC (0.76 β 0.78), which the authors interpret as much of the variation being explained by landscape factors rather than access alone; nevertheless, road/clinic proxies can still correlate with unmeasured behaviors and exposure.
Connections to related P. knowlesi evidence (external anchors)
The studyβs case-control design is tied to a previously published individual-level risk factor analysis in the same setting, indicating that exposure-related behaviors (e.g., walking to/from work or school) increase the risk of acquiring human P. knowlesi malaria.
The landscape variable construction relies on global forest cover change products from Hansen et al., which define forest canopy cover thresholds derived from Landsat data and are widely used in land cover change analysis.
Decision-usefulness: what this paper givesβand what it doesnβt
Gives: a practical framework for identifying which radii around households contain the strongest predictive signal from satellite-derived deforestation metrics using BRT RVI profiles.
Does not give: direct mechanistic mediation evidence showing whether the dominant effect operates through macaques, vectors, or human movement/behavior in this specific landscape.
Optional: run an AI Scientist agent to deepen this review
Author reviews (click to open BGPTβs bespoke reviews)
Feedback:
Updated: April 09, 2026
BGPT Paper Review
Study Novelty
90%
Novelty is high because the paperβs central contribution is a multi-scale inference framework for disease occurrence prediction that explicitly leverages BRT-derived scale signatures (RVI profiles) to constrain plausible mechanisms connecting deforestation-related landscape change to zoonotic malaria.
Scientific Quality
80%
Scientific quality is strong for study design and modeling transparency (case-control presence + surveyed absences, resampling to balance presences/absences, 10-fold cross-validation, explicit RVI computation) but constrained by inferential (not causal) interpretation and known risks in multi-scale correlational ML.
Study Generality
70%
Generality is fairly high because the scale-inference modeling approach can be reused for other zoonoses/vector-borne diseases; however, the specific predictor set (Hansen forest metrics, terrain, NDVI, access proxies) and the local ecology (hosts/vectors/macaque exposure) limit direct transferability of the exact 1 km and 4β5 km radii.
Study Usefulness
80%
Practical usefulness is high for surveillance prioritization and hypothesis refinement because it identifies scale peaks that can guide which neighborhood radii around households should be characterized and monitored when deforestation is changing.
Study Reproducibility
70%
Reproducibility is moderately strong: analyses are performed in R and code/environmental sample data are available on GitHub, but the human disease/household data are restricted via ethics procedures, which may limit full end-to-end replication for non-author users.
Explanatory Depth
80%
Explanatory depth is strong at the level of constraining mechanistic hypotheses via spatial scale signatures, while remaining weaker on direct causal mediation among macaques, vectors, and human behaviors.
None (not applicable): this review is primarily landscapeβdisease predictive modeling, not sequence- or proteomics-based bioinformatics on provided omics data.
Get emailed when your analysis is done!
We'll email you the results when your analysis is finished.
Hypothesis Graveyard
A βsingle buffer radiusβ mechanism hypothesis (e.g., only immediate β€1 km edge effects) is disfavored by the paperβs finding that fragmentation effects are strongest at larger radii (~4β5 km), implying at least two spatial scales with predictive signal rather than one uniform local mechanism.
A βvariance-only artifactβ explanation for all scale peaks is not fully supported because the authors report general agreement between RVI peaks and biological interpretation, while only some specific small-scale fragmentation/variance-linked cases require caution.
Explore by Goal
Quick Answer
Long Answer
Top Data Sources
Ask a Follow-Up
Key Insight
Keep Exploring
Analysis 
Hypothesis Graveyard
Potential Experiments
Discussion