Each project opens inside this portfolio as a complete case study: question, data, tools, workflow, analysis, visual evidence and interpretation.
A 53-participant behavioural study using E-Prime, JASP and Python to examine how state and trait anxiety relate to recognition accuracy across emotional valence.
Complete supervised-learning workflow on scikit-learn’s Breast Cancer Wisconsin Diagnostic dataset: inspection, preprocessing, EDA, PCA, train/test strategy, baseline, four candidate models, 5-fold cross-validation, evaluation, error analysis, explainability, SQL and BI implementation.
Testing whether self-oriented versus prosocial incentives change cognitive-task performance, then modelling the behavioural pattern computationally.
Eye-tracking analysis of where attention goes when anxious and non-anxious participants view emotionally charged imagery, with related PTSD work.
Behavioural analysis using correlation, ANOVA and post-hoc comparisons to test whether cognitive stress changes responses to emotional information.
A live 3 × 3 Noughts & Crosses reasoning model that evaluates every legal move and exposes why each machine decision was selected.
Pandas, NumPy and Matplotlib are evidenced in the end-to-end ML workflow and Recognition Bias visualisation pipeline.
scikit-learn, Logistic Regression, Decision Tree, Random Forest, KNN, PCA and cross-validation are implemented in the predictive-modelling case study.
Repeated-measures ANOVA, ANCOVA, Pearson correlation, Bonferroni comparisons and effect sizes are evidenced in the empirical Recognition Bias project.
State representation, candidate scoring and validation are demonstrated in the live Noughts & Crosses model.
SELECT, WHERE, GROUP BY, JOIN, CTE and window-function examples appear in the technical case study with explicit analytical purposes.
Power BI model structure and DAX measures are shown as a clearly labelled implementation example.
E-Prime, STAI, counterbalancing, eye tracking, AOIs, fixations, saccades and dwell time are tied directly to behavioural research case studies.
Reproducible classification workflow using scikit-learn’s Breast Cancer Wisconsin Diagnostic dataset. Results below are computed from that public dataset with an 80/20 stratified split and random_state = 42.
The technical objective is to inspect data quality, establish a baseline, compare model families under the same validation scheme, analyse errors and explain the final selection rather than optimise one isolated metric.
No imputation is required. The target is moderately imbalanced, so accuracy is supplemented by precision, recall, F1 and ROC-AUC.
Logistic Regression and KNN use standardised predictors. Decision Tree and Random Forest use the original numeric scale. All model selection is performed on the training data before final test evaluation.
PCA is used here for exploration rather than automatically replacing the original feature set in the final model.
| Model | CV accuracy | CV SD | Test accuracy | Precision | Recall | F1 | ROC-AUC |
|---|---|---|---|---|---|---|---|
| Logistic Regression | 97.8% | 1.1% | 98.2% | 98.6% | 98.6% | 98.6% | 99.5% |
| Random Forest | 96.5% | 2.0% | 94.7% | 95.8% | 95.8% | 95.8% | 99.4% |
| KNN | 96.3% | 2.3% | 97.4% | 96.0% | 100.0% | 98.0% | 98.8% |
| Decision Tree | 92.3% | 1.3% | 92.1% | 95.7% | 91.7% | 93.6% | 91.6% |
Logistic Regression produced the highest mean CV accuracy and 98.2% held-out accuracy. Its test confusion matrix contains one error in each off-diagonal cell.
Random Forest importances provide a model-specific view of influential predictors.
Filters high-perimeter cases for targeted review.
Compares class-level feature summaries before modelling.
Demonstrates relational enrichment with feature metadata.
Ranks observations within each target class using a CTE and window function.
Cases fact table + model predictions + feature metadata dimensions.
Logistic Regression is selected because it achieved the strongest cross-validated result and excellent held-out performance while remaining comparatively interpretable. Random Forest provides a useful nonlinear comparison and feature-importance view; KNN achieved perfect recall on this particular test split but slightly lower CV performance; the single Decision Tree underperformed.
A live 3 × 3 Noughts & Crosses model built to make decision scoring visible rather than opaque.
Noughts & Crosses is a constrained decision environment with nine positions and clear winning conditions. The goal was to build a model that does not merely choose a move; it also exposes why that move was preferred.
You are X. After every human move, the model scores the remaining legal squares and displays its reasoning trace.
A model demo is only useful if its behaviour can be checked. This validation suite runs deterministic board-state tests against the same scoring logic used by the live game.
Game-state reasoning, legal-action filtering, heuristic scoring, candidate ranking, interpretable decision traces, model validation, UI state management and implementation of a working computational model.
Testing whether self-oriented versus prosocial rewards change cognitive performance, then translating the observed behavioural pattern into a computational model.
The project compared self-oriented and prosocial/selfless incentive conditions to examine whether motivational framing changes cognitive-task performance.
This visual demonstrates the modelling concept only; it does not fabricate the original study’s empirical results.
Because the original participant-level dataset has not been supplied here, the panel below uses synthetic demonstration data. It proves the analytical workflow without presenting simulated values as empirical findings.
Analytical interpretation. A pattern like this would suggest that incentive framing is associated with performance differences, but the condition effect would need inferential testing before concluding that reward type reliably changes cognitive performance. A suitable workflow would estimate condition effects, confidence intervals and interaction terms, then test whether individual motivational parameters improve prediction beyond condition alone.
Published work supports the plausibility of reward-dependent cognitive effects while also showing that the effect can vary by reward type and participant characteristics. The portfolio therefore uses the literature to motivate the hypothetical pattern rather than presenting a single expected outcome as established fact.
Behavioural experimental logic, parameter estimation, computational replication, model fitting, model diagnostics and the ability to convert a psychological hypothesis into a quantitative modelling problem.
Examining where visual attention moves when anxious and non-anxious participants view emotionally charged images.
The project compared attentional behaviour in people with and without anxiety while viewing emotionally charged imagery. Related work also examined veterans with PTSD.
This demonstration shows how gaze transitions can be visualised without pretending to recreate participant-level values that were not supplied.
The table below is a synthetic eye-tracking example showing the type of derived dataset an employer would expect after preprocessing raw gaze events.
| Group | AOI | Fixations | Dwell time (ms) | Mean fixation (ms) |
|---|---|---|---|---|
| Anxiety | Threat AOI | 18 | 4,920 | 273 |
| Anxiety | Neutral AOI | 11 | 2,480 | 225 |
| Control | Threat AOI | 12 | 2,760 | 230 |
| Control | Neutral AOI | 13 | 3,120 | 240 |
| Illustrative measure | Lower anxiety | Higher anxiety | How it would be interpreted |
|---|---|---|---|
| First fixation to threat | 48% | 61% | Possible early vigilance |
| Latency to threat AOI | 620 ms | 505 ms | Faster orienting to threat |
| Threat dwell time | 1.18 s | 1.39 s | Greater sustained allocation |
| Neutral dwell time | 1.31 s | 1.20 s | Relative redistribution of gaze |
Why the time course matters. First fixation and latency address initial orienting; dwell time addresses later allocation or maintenance. Treating them as interchangeable can hide a vigilance–avoidance pattern in which attention is initially captured by threat but subsequently redirected away from it.
Eye-tracking research does not support one universal direction of effect. Adult studies have reported increased threat engagement under state anxiety, while other clinical and developmental work has found threat avoidance or time-dependent effects. A defensible interpretation therefore specifies which gaze metric, which stage of attention, which stimulus type and which population produced the effect.
Eye-tracking preprocessing, AOI definition, behavioural feature engineering, attention visualisation, group comparison and interpretation of spatial-temporal human data.
A 53-participant study examining how state and trait anxiety relate to recognition accuracy for negative, neutral and positive words.
Participants memorised emotionally valenced words and later classified each item as previously seen (“old”) or not previously seen (“new”). The analysis tested whether state and trait anxiety were associated with recognition accuracy across negative, neutral and positive conditions.
The first model tested recognition accuracy across valence and old/new status. State and trait anxiety were then included as covariates in ANCOVA. Bonferroni-adjusted comparisons were used where appropriate, and Pearson correlations examined anxiety–accuracy relationships within each condition.
Mean state anxiety was 49.58 (SD = 11.74), while mean trait anxiety was 52.32 (SD = 12.43). The report also classified participants into combined high/low state–trait profiles.
Bonferroni-adjusted comparisons showed that neutral/control-old words were recognised more accurately than control-new words were correctly rejected (p = .013). Positive-old accuracy was also higher than positive-new correct rejection (p < .001). The negative old–new difference was not statistically significant (p = .188).
The full matrix shows why a single overall “anxiety impairs memory” statement would be too simplistic. Associations change direction depending on emotional valence and whether the item is old or new.
State anxiety was only weakly and non-significantly related to control-new accuracy (r = .15, p = .29), but was negatively related to control-old accuracy (r = −.38, p = .006). Trait anxiety showed the same pattern: control-new r = .13, p = .342; control-old r = −.36, p = .009.
This mapping is a portfolio interpretation of the reported statistics. The key value is showing that hypotheses were evaluated against the observed direction of effects rather than retrofitted after analysis.
For previously seen words, mean accuracy was highest for positive items (M = .687) and lowest for negative items (M = .600). For new words, correct-rejection accuracy was highest for neutral items (M = .475) and lowest for positive items (M = .375).
Higher state anxiety was associated with better recognition of negative-old words (r = .73, p < .001) but poorer correct rejection of negative-new words (r = −.65, p < .001). Trait anxiety showed the same pattern: r = .67 for negative-old and r = −.66 for negative-new, both p < .001.
State anxiety correlated positively with positive-new accuracy (r = .52) and negatively with positive-old accuracy (r = −.51). Trait anxiety showed a similar pattern: r = .54 for positive-new and r = −.53 for positive-old.
The repeated-measures ANOVA found a significant valence × item-status interaction, F(2,104) = 3.50, p = .034, η²p = .063. When state and trait anxiety were included as covariates, the reported valence × item-status interaction became substantially larger, p < .001, η²p = .647. The three-way interaction with state anxiety was significant, F(2,100) = 4.201, p = .018, η²p = .078, whereas the corresponding three-way interaction with trait anxiety was not significant, F(2,100) = 2.260, p = .110, η²p = .043. This distinction suggests that momentary state anxiety contributed more strongly to the combined valence × familiarity pattern in that ANCOVA model.
Experimental design, behavioural data collection, psychometric measurement, repeated-measures analysis, ANCOVA, multiple-comparison control, correlation analysis, effect-size interpretation and Python visualisation.
Examining whether greater cognitive stress is associated with different behavioural responses to emotional information.
The project examined whether participants experiencing greater cognitive stress showed different behavioural responses to emotional information.
The portfolio presents the genuine analysis structure. Once the original output is supplied, this case study can be upgraded with the actual effect plots, group means, confidence intervals, test statistics and p-values.
The interactive example below uses synthetic demonstration data to show the difference between testing a continuous stress–response association and comparing mean responses across stress groups.
Hypothesis-driven analysis, data-quality checks, assumption testing, correlation, ANOVA, multiple-comparison follow-up and interpretation of behavioural effects.
Full résumé and supporting materials are available upon request.