RMD-Health: Machine learning-enabled decision support system to improve early detection and referral of rheumatic and musculoskeletal diseases

Background Over 20 million people in the UK live with rheumatic and musculoskeletal diseases (RMD) [5-6].

Inflammatory arthritis (IA) and non-inflammatory conditions (NIC) are two major subdivisions of RMD, each with very different treatment and management pathways (e.g. disease-modifying drugs for IA; surgeries such as joint replacements for NIC e.g. osteoarthritis).

Accurate and early detection and differentiation of IA and NIC are critical for patients to be referred to the right specialists and receive the right treatment rapidly.

Current machine learning (ML) research in RMD risk stratification relies on imaging examination [14-17] or specialised blood test results [18-19] conducted after referral to secondary care. They do not apply to early detection for RMD referral improvement in practice.

The existing triage automation solution [22] only predicts triage outcomes, meaning it is developed based on highly variable human referral triage assessments rather than confirmed diagnoses and cannot predict RMD diseases nor handle multimodal data.

We have developed the RMD risk stratification tool using data available at patients referrals and trained based on confirmed disease diagnoses.

Aim: This project aims to develop and validate an ML-based risk stratification decision support system, RMD-Health, for accurate and fast referral for patients with suspected RMD to a point where it can be submitted for regulatory approval compliant with AI as a Medical Device Regulations.

Objectives:   1) Evaluate the risk stratification model in large-scale retrospective data to detect model bias  2) Optimise and de-bias the risk stratification model to maximise accuracy and generalisability, and ensure fairness among patient subgroups through retrospective evaluation;   3) Develop a full software prototype of a real-time decision support system that incorporates user requirements;   4) Pilot and test the prototype in RBFT and OUH rheumatology referral pathways for usability, health economics, clinical safety and efficacy for regulatory approval and clinical adoption through prospective clinical validation.

Methods: Year 1: The RMD risk stratification model will be further evaluated and optimised on large retrospective rheumatology referral and diagnosis data.

The model will be debiased to ensure performance fairness across different patient groups with different demographic and socioeconomic including under-served groups.

Year 2: Software prototype will be co-developed and tested with clinical, PPI and NHS IT experts using an iterative agile approach. Year 3: The prototype will be piloted in RBH and OUH.

Clinical, operational and user feedback data from the prospective clinical validation will be used to finalise the prototype and technical file ready for regulation.

Impact: The RMD system will benefit 1) patients by enabling accurate and fast referrals thus reducing time waiting for the right treatment; 2) GP by reducing unnecessary consultations and workload; 3) secondary care clinicians by reducing unnecessary appointments and workload by automatic and more accurate risk stratification.

Our experiment shows it can reduce 8 hours per week (2 clinics) for referral assessment; 4) the NHS by reducing costs created by long referral times and unnecessary appointments.

This research will contribute directly to NHS long-term plan of transforming elective care including the delivery plan for tackling the COVID-19 backlogs.