Cases

The problem behind the method

These short cases focus on the decisions that made an analysis more defensible. They come from methodological work with students, but identities, named populations, specimens, collections and private results have been removed.

Problem → decision → outcome

What the analysis needed, not who supplied the data

My contribution in these projects was analytical: identifying a threat to the design, choosing or building a method, validating it and making the reasoning usable for the researcher whose question it was.

Data quality · shape analysis

When observer differences mimic a biological signal

Problem

An initial group comparison appeared interpretable, but measurements had been collected by different observers. A raw group effect could therefore reflect who measured the specimen rather than the structure being studied.

Solution

I redesigned the tests around the conditional question—group effect after accounting for observer—and controlled the resulting family of comparisons with false-discovery-rate correction. A landmark-participation table then localized where any remaining distance effect was concentrated.

What changed

The methodological threat became part of the model instead of a sentence in the limitations. The student could distinguish evidence about the research question from evidence about the measurement process.

Behind the analytical decision

Design checks

Test the research effect only after conditioning on observer identity.
Control the family of PC and distance tests instead of selecting an isolated low p-value.
Localise any surviving effect to contributing landmark pairs.

Validation

The revised analysis compared conditional group tests, false-discovery-rate-adjusted evidence and anatomical participation tables before interpreting a pattern.

Handover

A reproducible analysis sequence and a review table that separated measurement effects from biological interpretation.

Method design · geometric morphometrics

Adding curved anatomy without assuming it improves the answer

Problem

Fixed landmarks captured clear anatomical points but missed much of the curved form between them. Adding sliding semilandmarks could recover that information, yet it could also change the result simply by changing the representation.

Solution

I built the curve, redistribution and sliding workflow, then ran three parallel analyses: fixed landmarks only, semilandmarks only and both together. The design made the influence of the new representation visible instead of assuming that more points were automatically better.

What changed

The student received both a reusable curve-processing pipeline and a direct robustness check for the methodological choice at the centre of the project.

Behind the analytical decision

Design checks

Redistribute badly spaced curve points before alignment.
Keep fixed-only, curve-only and combined representations comparable.
Inspect deformation visualisations alongside ordination results.

Validation

Three matched analysis branches exposed whether the conclusion depended on adding and sliding curve semilandmarks.

Handover

A reusable curve-construction, redistribution and sliding pipeline plus presentation-ready shape comparisons.

Multivariate analysis · validation

Preventing one anatomical block from dominating another

Problem

Two linked anatomical regions had different landmark counts. Simply concatenating their coordinates would give the larger block more influence, regardless of its biological relevance.

Solution

I used a Multiple Factor Analysis-style balancing step before classification, then applied shrinkage discriminant analysis with stratified cross-validation and balanced accuracy. The held-out target remained completely outside training until the final classification.

What changed

Both anatomical blocks contributed on comparable terms, and the final probability could be interpreted as a genuinely held-out result rather than a consequence of leakage.

Behind the analytical decision

Design checks

Balance blocks before classification so landmark count does not become an implicit weight.
Stratify folds and use a class-balanced performance metric.
Keep the final target outside preprocessing, feature selection and training.

Validation

Cross-validation was completed on the reference data first; only then was the untouched target assigned posterior probabilities and distances.

Handover

A two-block analysis with outlier checks, classification diagnostics and visual shape reconstructions.

Surface analysis · feature engineering

Describing a rounded surface where landmarks are unreliable

Problem

The structure was approximately spherical and lacked enough repeatable anatomical landmarks for a conventional landmark-only analysis.

Solution

I introduced spherical harmonic decomposition as a surface descriptor, alongside rigid registration, PCA and age-related modelling. Scaled and unscaled routes ran in parallel so size effects remained separate from shape effects.

What changed

The analysis matched the geometry of the object instead of forcing the object into a familiar method, while preserving an interpretable comparison between size and shape.

Behind the analytical decision

Design checks

Rigidly align surfaces before extracting coefficients.
Run scaled and unscaled routes in parallel.
Compare harmonic descriptors with deformation- and landmark-based summaries.

Validation

Age trends and group separation were inspected across coefficient-space PCA, discriminant analysis and regression rather than from one model alone.

Handover

A surface-to-coefficient workflow with interoperable mesh conversion and parallel size-sensitive and size-free reports.

Clinical morphology · rule-based analysis

Turning anatomical expectations into reviewable screening rules

Problem

A large set of vertebral outlines needed consistent screening for unusual size progression and endplate shape. Manual review alone was slow and hard to reproduce.

Solution

I translated the anatomical expectation into geometry: polygon area, directional size progression and midpoint deviation from a reference line. Flags were severity-graded, reported in physical and relative units, and made less sensitive after false-positive review.

What changed

The workflow created a transparent shortlist for expert review rather than pretending to automate diagnosis. Sensitivity changes were explicit and measurable.

Behind the analytical decision

Design checks

Convert anatomical expectations into explicit geometric quantities.
Report deviations in both physical and relative units.
Tune thresholds after reviewing the false-positive pattern.

Validation

Flag counts and severity patterns were re-run after sensitivity changes, preserving a record of how the review threshold altered the shortlist.

Handover

A severity-graded workbook and plain-language report designed to support expert review, not replace it.

Classification · research design

Keeping a blind validation genuinely blind

Problem

A morphology classifier needed to be tested by independent observers without allowing target information to influence training, feature selection or score interpretation.

Solution

I separated development and blind-scoring stages, kept the targets outside the training workflow and evaluated performance with class-balanced metrics. Independent observer workbooks provided a real external test rather than a validation plan that existed only on paper.

What changed

The final result reflected performance on unseen scoring work, with the contribution of observer disagreement visible rather than absorbed into an optimistic headline metric.

Behind the analytical decision

Design checks

Freeze the development workflow before revealing blind targets.
Keep observer scoring independent of model fitting.
Use balanced metrics when classes are uneven.

Validation

Two independent scoring passes formed an external test set, allowing agreement and model performance to be evaluated separately.

Handover

A leakage-safe development pipeline, blind-scoring workbooks and an observer-aware validation report.

3D data · reusable infrastructure

Making irregular meshes comparable without a perfect template

Problem

Meshes arrived with inconsistent density, topology and local defects. Direct point-by-point comparison was impossible, while a manual repair-and-landmark workflow would not scale across projects.

Solution

I built a staged repair, alignment and equivalent-point pipeline with explicit quality checks and sampling options. The same infrastructure was handed on and appeared in several later research folders.

What changed

A one-off technical obstacle became shared infrastructure that other researchers could reuse and extend, reducing repeated preparation work across projects.

Behind the analytical decision

Design checks

Repair only the defects that block correspondence while retaining biological geometry.
Separate rigid alignment from equivalent-point sampling.
Expose sampling and quality-control choices instead of burying them in defaults.

Validation

Intermediate meshes, alignment residuals and point distributions were inspected at each stage before results were handed downstream.

Handover

A staged, configurable surface pipeline that was reused and extended in later student and collaborative work.

Predictive modelling · uncertainty

Estimating from incomplete measurements without hiding uncertainty

Problem

Age estimation had to combine a small reference sample, repeated measurements from different observers and variables that were not equally reliable for interpolation.

Solution

I designed a staged pipeline that discovered candidate variables inside grouped validation, modelled observer error and down-weighted measurements whose target value lay outside a defensible reference range.

What changed

The estimate carried information about measurement reliability and interpolation safety instead of presenting one precise number unsupported by the available reference data.

Behind the analytical decision

Design checks

Discover candidate variables inside grouped validation folds.
Estimate observer error from replicate measurements.
Reduce the influence of variables that require unsafe extrapolation.

Validation

The pipeline kept replicate observations grouped, repeated variable selection within validation and reported which inputs were reliable enough to contribute.

Handover

An observer-aware estimation workflow with variable weights, interpolation flags and uncertainty visible in the final report.