This project was completed as part of an industry placement at the University of Southern Queensland (USQ) through the CSC6205 Health Informatics unit. It examines healthcare fragmentation and service gaps across four rural LGAs in South-West Queensland, combining descriptive analytics, predictive modelling, and unsupervised clustering.
applied-analytics-health-infomatics/
├── health_informatics_analysis.ipynb ← Main combined analysis notebook
├── README.md
├── .gitignore
├── AA_Assignment_1/ ← Source data + reports (incl. the input
│ └── ... 'finalised dataset.csv' read by the notebook)
├── AA_Assignment_2/ ← Maranoa postcode dataset + reports
│ └── ...
├── figures/ ← All generated plots (19 PNG files)
│ ├── fig_01_health_conditions_by_lga.png
│ └── ...
└── outputs/ ← All exported data tables (13 CSV files)
├── descriptive_health_stats.csv
└── ...
Python 3.10
├── pandas — data manipulation and descriptive statistics
├── numpy — numerical operations
├── matplotlib — base visualisation
├── seaborn — statistical visualisation
├── scikit-learn — StandardScaler · LinearRegression · KMeans · PCA
└── scipy.stats — spearmanr · kruskal · chi2_contingency
# Clone the repository
git clone <repo-url>
cd applied-analytics-health-infomatics
# Install dependencies
pip install pandas numpy matplotlib seaborn scikit-learn scipy jupyter
# Launch the notebook
jupyter notebook health_informatics_analysis.ipynbThe notebook is self-contained and runs top-to-bottom. All figures are written to figures/ and all CSV outputs to outputs/.
Four LGAs in South-West Queensland were compared using integrated ABS Census, NDIS participant, and National Health Survey 2022 data (540 records × 116 variables after cleaning; each record is a postcode × age-group × sex aggregate).
- Removed aggregate
Totalrows - Imputed numeric missing values with column median; categorical with mode
- Standardised LGA names to lowercase
Arthritis and asthma are the most prevalent conditions across all LGAs. Western Downs shows higher absolute counts proportional to population, but prevalence rates are comparable or higher in the more remote LGAs. Mental health conditions are notable across all four regions.
Western Downs has the highest NDIS participation (~102 participants and ~148 providers per record on average). Quilpie has the lowest (~9 participants, ~9 providers), suggesting major access gaps in the most remote LGA.
Healthcare workforce is heavily concentrated in Western Downs (inner-regional). Maranoa, Murweh, and Quilpie report near-zero employees in specialist medical, pathology, and allied health categories — confirming systemic workforce fragmentation in remote areas.
Indigenous representation is highest in Murweh and Quilpie, the most remote LGAs. This intersects with documented disparities in healthcare access and health outcomes for Aboriginal and Torres Strait Islander communities in remote regions.
Strong positive correlations are observed between most chronic conditions, particularly arthritis–diabetes, arthritis–heart disease, and arthritis–kidney disease. This co-morbidity clustering is consistent with the population being predominantly older and rurally isolated.
Box plots reveal high within-LGA variance, particularly for Western Downs, reflecting its greater postcode-level heterogeneity. Quilpie and Murweh show tighter distributions consistent with their smaller, more homogeneous populations.
Thematic descriptive statistics (min · Q1 · median · mean · Q3 · max · std) were generated for six domains and exported to outputs/:
| File | Domain |
|---|---|
outputs/descriptive_health_stats.csv |
Chronic health condition prevalence |
outputs/descriptive_ndis_stats.csv |
NDIS participant and provider counts |
outputs/descriptive_employment_stats.csv |
Industry employment by LGA |
outputs/descriptive_healthcare_stats.csv |
Healthcare sector employment by LGA |
outputs/descriptive_migration_stats.csv |
5-year residential migration patterns |
outputs/descriptive_assistance_stats.csv |
Need for assistance and unpaid care |
Target variable: has_need_for_assistance_with_core_activities
Features were standardised with StandardScaler, age encoded ordinally, and sex/LGA one-hot encoded. An 80/20 train/test split was applied (384/96 rows, 111 features).
| Metric | Value |
|---|---|
| R² | 0.570 |
| RMSE | 6.92 |
| MSE | 47.85 |
The target is a count (people needing core-activity assistance per record), so the RMSE of ~6.9 is in persons per postcode × age × sex cell.
The strongest predictors of assistance need are household relationship variables (particularly child presence in the home) and complementary assistance indicators. employed_arts_recreation_services appears as a positive predictor, likely a proxy for community connectedness and social infrastructure in better-resourced areas.
ANOVA confirms that assistance-related variables and household composition features are the most statistically significant predictors. Features with p > 0.05 (grey) contribute little independent predictive power once other variables are accounted for.
Four hypothetical policy interventions were simulated against the trained model (mean change in predicted assistance need):
| Intervention | Mean Δ predicted need | Interpretation |
|---|---|---|
| Aged care service presence (set to 1) | +0.86 | Increase — aged care presence correlates with concentrations of higher-need populations |
| NDIS providers +20% | +0.01 | Near-zero effect |
| Health employment +20% | +0.02 | Near-zero effect |
| Hospital presence (set to 1) | −0.19 | Modest reduction in predicted need |
The near-zero NDIS provider effect is consistent with the report findings — increasing provider numbers does not help when geographic accessibility barriers remain unaddressed. These simulations probe a correlational model, so the shifts indicate association rather than causal policy effect.
The Maranoa Regional Council was profiled at postcode level (n = 10, 210 variables) using unsupervised clustering to identify distinct community subregions.
Roma is by far the largest settlement (population 8,610) and the highest-income suburb ($725/week median). Muckadilla and Eurella are at the far end of both measures, with populations under 50.
The elbow at k = 3 indicates three distinct community profiles within Maranoa, which maps intuitively onto the geographic and demographic structure of the region.
PCA reduces the 200+ variable space to two principal components. Cluster 1 (Roma) is clearly separated from the others — representing the regional hub. Cluster 0 contains small, very-remote settlements; Cluster 2 contains moderate-sized remote communities. Bubble size is proportional to total population.
| Cluster | Population | Median Income ($/wk) | Avg Income ($/wk) | Practitioners | Ambulance Time (min) |
|---|---|---|---|---|---|
| 0 — Remote small | 235 | 587 | 945 | 1.8 | 79.8 |
| 1 — Roma hub | 8,610 | 725 | 1,131 | 50 | 109 |
| 2 — Mid-size remote | 897 | 583 | 976 | 29 | 101 |
Cluster 1 (Roma) holds the vast majority of every practitioner type. Cluster 0 has effectively no specialist coverage. Cluster 2 has some coverage but far below population-proportionate levels.
Absolute condition counts are highest in Cluster 1 simply due to population size. However, prevalence rates per capita are comparable or higher in Clusters 0 and 2 — smaller communities bear proportionally similar disease burden with far less healthcare capacity.
| Condition | ρ | p-value | Significant |
|---|---|---|---|
| Diabetes | 0.891 | 0.0005 | yes |
| Asthma | 0.816 | 0.004 | yes |
| Lung condition | 0.773 | 0.009 | yes |
| Heart disease | 0.762 | 0.010 | yes |
| Mental health | 0.723 | 0.018 | yes |
| Arthritis | 0.717 | 0.020 | yes |
| Stroke | 0.691 | 0.027 | yes |
Seven of the ten tested conditions show significant positive correlation with cluster (dementia, cancer, and kidney disease do not reach significance). Communities with more concentrated healthcare (Cluster 1) also have higher recorded condition counts, driven primarily by population size and better detection/recording rates.
Nine of the twelve practitioner types show statistically significant variation across clusters (p < 0.05), suggesting the clustering captures real differences in healthcare workforce distribution rather than random variation — though with only 10 observations, these tests are indicative rather than conclusive.
Five employment sectors differ significantly across clusters — health care & social assistance, manufacturing, retail, accommodation & food services, and public administration & safety (education & training and mining do not reach significance) — reflecting the broader socioeconomic differences between the Roma hub and surrounding remote communities.
Cluster 1 (Roma) holds the only hospital with advanced capabilities. Clusters 0 and 2 have no recorded capacity for cardiac treatment, intensive care, neonatal care, nuclear medicine, or surgical oncology.
-
Healthcare workforce is heavily concentrated in the least remote LGA. Western Downs holds the bulk of healthcare-sector employment, while Maranoa, Murweh, and Quilpie report near-zero employees in specialist medical, pathology, and allied health categories.
-
Chronic conditions co-occur. Arthritis, diabetes, heart disease, and kidney disease are strongly correlated across records, consistent with an older, rurally isolated population carrying multiple conditions at once.
-
Assistance need is predictable from census aggregates (R² = 0.57), driven primarily by household composition and complementary assistance indicators — though the inclusion of the complementary
does_not_have_need_for_assistancevariable inflates this figure (see Limitations). -
Simulated service-level interventions barely move predicted need. Increasing NDIS providers or health employment by 20% changes predicted assistance need by ≈0; only hospital presence shows a (modest) reduction. This is consistent with the report's conclusion that funding-only levers don't resolve geographic access barriers, but as the model is correlational these results should not be read as causal policy estimates.
-
Maranoa has three structurally distinct subregions. Roma (Cluster 1) is the clear healthcare hub, holding the large majority of every practitioner type. Cluster 0's small remote settlements average only 1–2 registered practitioners each, so a single departure can collapse local coverage entirely.
-
Critical clinical service gaps are systemic. Intensive care, nuclear medicine, and surgical oncology are absent from all Maranoa postcodes, forcing high-acuity patients to travel hundreds of kilometres to tertiary facilities.
| Dataset | Source |
|---|---|
| ABS Census 2021 (age, sex, household, employment, migration, background) | Australian Bureau of Statistics |
| NDIS Participants & Providers by Postcode | NDIS Data Explorer |
| National Health Survey 2022 | ABS |
| Healthcare Workforce Shortage Data | Australian Government DoHAC |
| Hospital Clinical Services Capability | South West Hospital & Health Service |
| Practitioner Registrations by Postcode | AHPRA |
| Charity Registry | ACNC |
| Gross Regional Product | .id Consulting / Economy.id |
| Hospital Annual Report 2023–24 | South West Hospital & Health Service |
- The Assignment 1 dataset is aggregated at postcode × age-group × sex level — individual-level inference is not valid
- Assignment 2 has only 10 postcode-level observations; statistical tests should be interpreted with caution given the small sample size
- The linear regression model includes a complementary variable (
does_not_have_need_for_assistance) that inflates predictive performance — feature selection refinement is recommended for production use - Cross-sectional data cannot establish causality
Completed as part of an industry placement at the University of Southern Queensland (USQ), CSC6205 Health Informatics, Applied Analytics stream.