Software Sushi JournalHealthcare workflow pattern
A Risk Estimate Is Not the Decision
Keep the model-generated estimate and the professional's recorded judgment as different records in the same workflow.

Logical workflow based on project documentation
How the responsibilities connect
Step through the documented path. These are logical responsibilities, not a diagram of a particular cloud deployment or technology stack. The illustration is conceptual and contains no patient records.
A response can reveal a dependent question
The assessment interface collects structured responses. Conditional questions and completion checks apply before the user proceeds. Hidden branches stay out of the path.

A risk estimate is not the end of a healthcare assessment. It has to reach the professional who reviews the case, stay connected to the information collected, and remain distinct from the decision that person records.
Software Sushi delivered this pattern in a healthcare project that reached production: risk-prediction models and a prediction service, a structured assessment interface, and backend review workflows. The public write-up of that work is anonymous. It includes no patient records, customer identifiers, or application screenshots. This article explains the same pattern, based only on that healthcare risk prediction case. The capabilities described reflect project documentation, not an audited inventory of every deployed version.
A model estimate is not the recorded decision
A common failure in a prediction product is to treat the model's response as the end of the interaction. In this workflow, that is the wrong boundary. The documented design brings model-generated results into professional review alongside the assessment, rather than treating a prediction response as the completed interaction.
The backend supports explicit operations for recording clinical decisions and interventions. That is how the product preserves the distinction: what the model estimated and what the professional decided remain different records.
Structured assessments come first
The assessment interface collects the information that later moves through the workflow. We built that interface with conditional questions. A response can reveal a dependent question, while other answers leave that branch hidden. Completion checks apply to visible follow-up questions before the user proceeds.
Those interactions connect the structure of the questionnaire to the information passed to prediction and review. Hidden branches stay out of the path. Visible follow-ups have to be complete.
Give prediction a service boundary
We built risk-prediction models and exposed prediction functionality through a dedicated API. The services accept questionnaire data and return risk-related estimates for use in the assessment workflow.
That service boundary gives prediction a defined role in the application. The assessment interface collects information. The prediction service produces an estimate. The review workflow lets a professional consider the result and record a decision. The API stops at the estimate.
Bring the estimate into review, with context
The estimate should not appear on its own. The documented workflow brings model-generated results into professional review alongside the assessment.
Our backend work covered assessment records, history, and clinical-review operations. Project documentation also describes configurable follow-up frequencies and aggregate assessment views, which extend the application beyond a single submission and result.
The point is a connected review process: collect the assessment, bring the estimate into view, and give professional judgment an explicit place in the application.
Record the clinical decision as its own operation
Review is not complete when the professional has seen the estimate. The backend supports explicit operations for recording clinical decisions and interventions. The recorded decision is written as its own piece of information, not as a replacement for the estimate.
That recording step keeps the distinction in the product: the estimate and the decision remain different pieces of information.
How the responsibilities connect
The case describes a logical workflow, based on project documentation:
- Structured assessment
- Prediction service
- Model-generated risk estimate
- Professional review
- Recorded clinical decision
Three groups of responsibility sit on that path. Prediction models and services produce risk-related estimates from questionnaire data and expose them to the application. The assessment interface collects structured responses and handles conditional questions and completion checks. Backend review workflows manage assessment records, history, clinical-decision updates, and follow-up settings.
These are logical responsibilities, not a diagram of a particular cloud deployment or technology stack.
Production delivery is not clinical validation
Software Sushi's contribution spanned prediction modeling and services, the assessment experience, and backend review workflows. The project reached production. The engineering result is an application around the prediction, not just an isolated model endpoint. It connects the information being assessed, the estimate being reviewed, and the judgment being recorded.
That production delivery has a clear clinical limit. Risk estimates are inputs for qualified professional review, not autonomous diagnosis or treatment instructions. Production delivery alone does not establish clinical validation, regulatory approval, compliance certification, model accuracy, or improved patient outcomes.
The public case, and this article, stay inside those limits. The capabilities described reflect project documentation, not an audited inventory of every deployed version. The case is anonymous and includes no patient records, customer identifiers, or actual application screenshots. Read the healthcare risk prediction case for the same source narrative.
Frequently Asked Questions
Is a model-generated risk estimate a clinical decision?
No. In this workflow, the estimate and the recorded decision are different pieces of information. A prediction response is not the completed interaction. A qualified professional reviews the estimate alongside the assessment and records a decision through an explicit operation.
What does the prediction service return?
The services accept questionnaire data and return risk-related estimates for use in the assessment workflow. They do not record the clinical decision.
Does reaching production mean the model is clinically validated?
No. Production delivery alone does not establish clinical validation, regulatory approval, compliance certification, model accuracy, or improved patient outcomes.
Does this article describe a named client or the live application?
No. The case is anonymous and includes no patient records, customer identifiers, or actual application screenshots.
Bring a prediction capability into a usable workflow
The useful unit of delivery in this work was not an isolated model endpoint. It was the application around the prediction: structured collection, a defined service boundary, review with context, and a recorded professional decision.
If you are building a healthcare assessment or decision-support product, we can help connect prediction engineering with the interfaces, application services, and review steps around it. For a similar workflow, explore a fixed-scope healthcare AI pilot, or book a call.


