Critical Preliminaries

Components

	Notes
Problem Statement	What is the problem, and why is it important to solve this problem?
Outcome Expectations, Underlying Aims	What are the potential product's outcome expectations, e.g., An increase in the detection of early-stage, i.e. mild, diabetic retinopathy, hence the prevention of sight loss amongst a greater percentage of diabetic patients. The underlying aim might be, e.g., To reduce the mis-classification rate of retinal images of diabetic patients to ≤ 1.25% per class per quarter.
Deployment Goal	What is the aim, expected output, of the product's model? Example: A retinal image's diabetic retinopathy classification probability per class. What will the deployed product present to the optician? Or rather, how will the model's output be used? Considering the example above, the client's opticians might expect: The classification probability per class. The definition of each class. The classification; vis-à-vis the highest classification probability. The preliminary treatment pathway of the classification, according to the rules of the optician's institution. The retinal image.

Table Footnote:¹

In Practice

An example of a problem statement, outcome expectations & underlying aims, and deployment goal.

Problem Statement

An organisation manually classifies trauma incidents for all the major trauma centres of five countries. Per trauma case, an injury coding expert (a) examines the case's free and structured text, and assigns each piece of text to a category, and (b) assigns the case to a trauma category based on the combination text pieces & categories detected; text pieces of the other/miscellaneous category are excluded from this exercise. Trauma injury coding is an extremely intensive and time-consuming exercise, and injury coding error rates - per annum - can be quite high. Hence, and as a first step, we are in search of a solution that automatically classifies text pieces vis-à-vis a set of provided categories.

Outcome Expectations, Underlying Aims

• Outcome Expectations: Real-time availability of classifications per trauma case.
• Underlying Aim: The automatic classification of trauma case text pieces; objective →
      - Per case, automatic classification time < 180 seconds.
      - Model performance per class → false negative rate ≤ 0.02, false positive rate ≤ 0.04

Deployment Goal

A potential machine learning dependent project without a deployment goal is directionless, do not proceed. A plausible deployment goal is

---

1. This is not machine learning. This is about integrating a machine learning model's output into a business product, and hence business operations. ↩