Scenario

• Growing ICT impact on clinical practice and research
  • ML >> predictive modelling >> decision support
  • i.e. risk prediction
• Barriers to overcome
  • standards
  • technological development

Use cases

1. Data science teams want to join efforts for better predictive models
   • geographical distribution >> remote collaboration only
   • different organisations >> data sharing cumbersome
   • different technical skills >> sharing code barely useful

2. Clinical staff wants to experiment predictive modeling
   • lacks IT resources >> up-front investment
   • lacks IT skills >> steep learning curve

3. Healthcare organisations want to assess transferability of models
   • lack reciprocal trust >> no data disclosure
   • different technical skills >> incompatible models

Proposition

Ubiquitous access & seamless deployment

of prediction models.

Interoperability & collaboration

between data science teams and, possibly, clinicians.

Architecture

Web API: RESTful endpoints

Translation: R/Python models to PMML/PFA

Prediction: Serve predictions

Learning: Build prediction models

Storage: Store models

Functionalities

• Common:
• set global (usable anywhere vs. provider site)
• set online (learning on new data vs. "frozen")

Non-functional Properties

Trust: local validation + transparency (inspection)

Language / framework agnostic: automatic translation + PMML/PFA (R, caret + Python, scikit-learn)

Interoperability: PMML/PFA, HTTP API, JSON/XML data

Accessibility: Web Service, REST architecture

Enabling technologies

• "X-as-a-Service" paradigm
• especially: "ML-as-a-Service" (MLaaS)
• Standards promoting seamless exchange and deployment of ML models
• don't brake the data pipeline
• language / framework agnostic >> "build once, run everywhere"

Service Orientation

• "as-a-Service" paradigm to deliver software products
• roots in Service Oriented Architecture (~90s)
• well-known implementation: Web Services (~2000s)
• modern evolution: RESTful micro-services (~2010s)
• MLaaS growing trend
• i.e. IBM Watson ML, Amazon SageMaker, MS Azure ML Studio, Google Cloud ML Engine
• off-the-shelf solutions : wide portfolio of models, customisation options, user friendly
• lack focus on serving models (i.e. IBM MAX)

PMML / PFA

independent development / deployment of predictive models

flexibility of production environment

co-existence of different programming languages and frameworks

Application

• Architecture implemented in CONNECARE European project
• Predictions about hospitalisations, emergency re-admissions, mortality
• Models created by partners with locally available data (6000+ patients with 5+ years follow-ups)
• DSS as precious asset to let clinical partners
  • use and validate each other models
  • cooperatively improve models without data sharing
  • assess transferability

Risk Prediction

Home Hospitalisation (HH) / Early Discharge (ED)

Conclusion

Reference architecture for a DSS, promoting

State of art tech + ML standards = support to practical use cases

Running implementation within CONNECARE

Thanks

for your attention

Questions?

Stefano Mariani

Università di Modena e Reggio Emilia