A 4-Layer Integrity Intelligence System

Each layer adds a dimension of intelligence -- from raw relationship mapping through temporal modeling, physical validation, and graph-based risk scoring. Together they produce a continuously updated integrity signal that scales with the NDIS.

Layer 1: Care Graph Layer 2: Behaviour Layer 3: Feasibility Layer 4: AI Risk
1

Care Network Graph

Multi-Relational Dynamic Graph

The foundation of CareIntegrity.AI. Instead of storing claims as flat rows in a database, Layer 1 builds a continuously updating knowledge graph of the entire care economy -- mapping every provider, participant, worker, location, and the relationships between them.

Graph Entity Model

Provider
Worker
Participant
Provider
Location
Provider

Edges represent: service delivery, billing, employment, co-location, shared staffing

5
Entity Types
6
Relationship Types
550+
Active Nodes
10K+
Edges

Graph Entities (Nodes)

  • Providers -- registered NDIS service organisations
  • Participants -- individuals receiving NDIS support
  • Workers -- support workers and allied health professionals
  • Locations -- SIL houses, offices, community venues
  • Services -- specific NDIS service categories

Relationship Edges

  • Bills -- provider invoices participant for services
  • Employs -- provider engages worker
  • Serves -- worker delivers care to participant
  • Operates At -- provider operates from location
  • Registered With -- worker registered across providers
  • Co-attendance -- shared service delivery patterns

What the Graph Reveals

Provider A and Provider B appear independent in a relational database. But the Care Graph shows they share 6 workers, operate from the same address, and bill the same 30 participants. They are operationally one entity hiding behind two registrations -- a pattern invisible without graph analysis.

Technical Implementation

Built on NetworkX with directed weighted edges. Node attributes include geographic coordinates, registration data, and entity metadata. Edge weights represent billing volume, service frequency, and employment strength. The graph updates continuously as new claims data arrives, maintaining a real-time representation of the care economy.

2

Temporal Behaviour Engine

Behavioural Fingerprint Evolution Over Time

Layer 2 adds the dimension of time. It doesn't just detect anomalies -- it detects drift. Every provider's behaviour is modeled as a monthly fingerprint across 15+ dimensions. When these fingerprints change faster than any legitimate business evolution could explain, the engine identifies structural impossibility.

Provider Fingerprint Over Time

Jan
Feb
Mar
Apr
May
Jun

Sudden billing spike after months of stable behaviour = impossible acceleration signal

Billing Patterns

  • Avg hours per participant
  • Total monthly billing
  • Rate per hour trends
  • Session length distribution

Workforce Metrics

  • Participant-to-worker ratio
  • Worker count growth
  • Staff turnover patterns
  • Multi-provider workers

Temporal Signals

  • Peak billing hour shifts
  • Weekend ratio changes
  • Geographic spread expansion
  • Service mix evolution

Impossible Acceleration Example

Provider PRV-0007 had 5 participants and 2 workers in January. By March, they had 60 participants -- but still only 2 workers. That's a 30:1 participant-to-worker ratio, up from 2.5:1. No legitimate provider can serve 30 participants per worker. The Temporal Engine flags this as structurally impossible -- the billing is fabricated.

Service Mix Drift Example

A registered Occupational Therapy provider's monthly fingerprint shows their service mix shifting from 90% OT in January to 15% OT / 70% SIL by June. Legitimate providers don't fundamentally change their service category within 5 months. This semantic role shift indicates either a takeover or opportunistic billing fraud.

3

Care Feasibility Validator

Physics-Based Constraint Engine

Layer 3 asks the most fundamental question: "Is this care delivery physically possible?" It enforces the immutable laws of physics -- a person can only be in one place at a time, a day has exactly 24 hours, and travel between locations takes time. Claims that violate these constraints are fabricated, period.

Three Constraint Models

T
Time
Cannot bill more than 24h in a day
S
Space
Cannot be in two places at once
D
Distance
Travel takes measurable time
100
Time Impossibilities
50
Excessive Daily Hours
50
Travel Impossibilities
100
Overservicing Flags

Worker Constraints

  • Simultaneity: overlapping sessions at different locations
  • Daily capacity: total hours exceeding 16h/day
  • Travel feasibility: insufficient gap for distance between sessions
  • Weekly limits: exceeding 38h/week across all providers

Participant Constraints

  • Plan allocation: hours exceeding weekly entitlement
  • Budget limits: spending approaching or exceeding plan budget
  • Service frequency: sessions beyond clinical justification
  • Concurrent services: too many therapy types simultaneously

Simultaneous Billing Example

Worker WRK-0005 is billed for a session in Parramatta from 10:00-13:00 and simultaneously billed for a session in Penrith from 10:30-14:00. These locations are 25km apart. The Feasibility Validator flags this instantly -- one of these sessions is fabricated. The worker cannot physically be in both places.

Travel Impossibility Example

Worker WRK-0013 finishes a session in Liverpool at 14:00 and starts another session in Hornsby at 14:10. These suburbs are 55km apart by road. At 60km/h average urban speed, this journey takes at least 55 minutes. A 10-minute gap is physically impossible -- the Hornsby session is fabricated or the times are falsified.

4

Graph AI Risk Engine

Machine Learning on Graph Structures

Layer 4 is where artificial intelligence meets graph theory. Every provider is converted into a high-dimensional behaviour vector (embedding). Graph Neural Network techniques then detect patterns that no human analyst and no rule-based system could find -- collusive clusters, synthetic behaviour patterns, and abnormal network centrality shifts.

From Raw Data to Risk Intelligence

Claims Data
Feature Extraction
35-Dim Embedding
PCA / Clustering
Risk Score
35
Embedding Dimensions
60
Provider Vectors
30
Mutations Detected
170
Cluster Anomalies

Technology Stack

  • Graph Neural Networks (GNN) architecture
  • Node embeddings (provider behaviour vectors)
  • Unsupervised anomaly detection
  • Community detection (Louvain / spectral clustering)
  • PCA with force-spread for 2D visualisation
  • Cosine and Euclidean distance metrics

What It Detects

  • Collusive provider clusters (shared infrastructure)
  • Abnormal network centrality shifts
  • Synthetic-looking provider behaviour patterns
  • Behavioural mutations (sudden DNA shifts)
  • Providers clustering with known fraud entities
  • Emergent fraud patterns invisible to rules

Embedding Features (35 dimensions)

  • Service mix proportions (12 NDIS categories)
  • Participant demographics served (7 disability types)
  • Session duration and rate statistics
  • Weekend and night billing ratios
  • Geographic spread and billing intensity
  • Time distribution entropy (8 bins)
  • Rate and duration variance measures

Risk Intelligence Outputs

  • Provider Risk Score: 0-100 composite across all layers
  • Network Cluster Risk Index: community-level assessment
  • Behavioural Drift Score: temporal deviation from baseline
  • Constraint Violation Events: real-time impossibility flags

Integrates via real-time API or batch pipelines -- no rip-and-replace required.

Cluster Anomaly Example

The Graph AI Risk Engine identifies that Provider PRV-0023 has a cosine similarity of 0.92 with known fraud provider PRV-0003. Their embeddings are nearly identical -- same service mix, same billing patterns, same time-of-day distribution, same geographic spread. PRV-0023 was not previously flagged, but its behaviour DNA is a near-perfect match for a confirmed fraudulent entity. This triggers automatic investigation.

See All 4 Layers Working Together

Live demo with real detection across all layers -- from graph construction to risk scoring.

Access Live Demo
Login: admin / NDISAdmin2025!