JMIR Publications

ABSTRACT

Background:

Electronic Medical Records (EMRs) are increasingly used to support machine learning (ML) in healthcare, yet comparatively little attention has been given to the translational work required to prepare routinely collected operational data before modelling can occur.

Objective:

This study examined the technical, clinical, and operational work required to prepare EMR data for downstream ML applications and developed an interprofessional translational preprocessing framework.

Methods:

An observational study was undertaken using routinely collected EMR data from 398,872 inpatient admissions at a South Australian quaternary hospital between 2020 and 2025. The preprocessing methodology integrated technical data preparation with workflow interpretation, interprofessional validation, and feature derivation to develop analytically meaningful and operationally interpretable model inputs. The study is reported in accordance with the RECORD guidelines.

Results:

Preprocessing identified recurrent challenges extending beyond conventional data cleaning, including incomplete structured documentation, workflow artefacts, high-cardinality variables, free-text dependence, and temporal instability. Many variables initially assumed to represent straightforward clinical or operational concepts instead reflected documentation practices, workflow behaviours, or administrative processes, requiring interprofessional interpretation before they could be reliably used for ML. These recurring activities informed development of a six-stage interprofessional translational framework spanning shared problem definition, data validity assessment, workflow interpretation, interprofessional validation, feature translation, and translation into action.

Conclusions:

Preparing routinely collected EMR data for ML is not solely a technical exercise in data cleaning but an interprofessional translational process. The framework described provides a structured approach for preparing operational healthcare data for ML applications while preserving clinically meaningful and operationally interpretable representations of healthcare activity. Health systems seeking to realise value from AI should invest not only in technical modelling capability, but also in the organisational capability required to steward, interpret, validate, and translate routinely collected healthcare data into actionable intelligence.