Using artificial intelligence to predict patient outcomes from patient-reported outcome measures: a scoping review

Abstract

Purpose: This scoping review aims to identify and summarise artificial intelligence (AI) methods applied to patient-reported outcome measures (PROMs) for prediction of patient outcomes, such as survival, quality of life, or treatment decisions.

Introduction: AI models have been successfully applied to predict outcomes for patients using mainly clinically focused data. However, systematic guidance for utilising AI and PROMs for patient outcome predictions is lacking. This leads to inconsistency of model development and evaluation, limited practical implications, and poor translation to clinical practice.

Materials and methods: This review was conducted across Web of Science, IEEE Xplore, ACM, Digital Library, Cochrane Central Register of Controlled Trials, Medline and Embase databases. Adapted search terms identified published research using AI models with patient-reported data for outcome predictions. Papers using PROMs data as input variables in AI models for prediction of patient outcomes were included.

Results: Three thousand and seventy-seven records were screened, 94 of which were included in the analysis. AI models applied to PROMs data for outcome predictions are most commonly used in orthopaedics and oncology. Poor reporting of model hyperparameters and inconsistent techniques of handling class imbalance and missingness in data were found. The absence of external model validation, participants' ethnicity information and stakeholders involvement was common.

Conclusion: The results highlight inconsistencies in conducting and reporting of AI research involving PROMs in patients' outcomes predictions, which reduces the reproducibility of the studies. Recommendations for external validation and stakeholders' involvement are given to increase the opportunities for applying AI models in clinical practice.

Fig. 1

PRISMA flow diagram

Fig. 2

Year of publication of all 94 studies (top figure) and studies based on health domain (bottom figure)

Fig. 3

Number of papers categorised based on sample sizes in each healthcare domain

Fig. 4

Reporting of pre-processing and model development methods in the studies. Sub-figure a) Frequency of hyperparameter tuning and values reporting. Sub-figure b) Proportion of hyperparameter tuning techniques. Sub-figure c) Missingness reporting and imputation in papers. Sub-figure d) Handling class imbalance in studies

Fig. 5

Frequency of algorithms used on datasets with very small (fewer than 300), small (300–700), medium (701–2000), large (2001–20000) and very large (more than 20,000) sample size