Artificial Intelligence and Patient-Reported Outcome Measures for Predicting Hospital Utilisation and Health-Related Quality of Life During Chemotherapy

Patient-reported outcome measures (PROMs) are any reports provided by the patient without interpretation from a clinician or anybody else. The use of PROMs in clinical practice can facilitate shared decision-making and self-management of symptoms. PROMs are increasingly being used in Artificial Intelligence (AI) models to predict patient outcomes. However, the lack of guidance on using PROMs in AI models leads to inconsistencies in data pre-processing, model development and model evaluation. This hinders the practical adoption of PROMs and AI models and limits their potential to improve patient outcomes.
Through a literature scoping review, the following research gaps were found: 1) Unclear predictive value of PROMs; 2) Inconsistent data pre-processing methods; 3) Lack of patient and clinician involvement in the study design; 4) Unexplored potentials of time-series PROMs and deep learning (DL) models. To address these gaps, this thesis developed a novel, patient-centric framework for applying AI methods on PROMs. Data collected in eRAPID clinical trial were used to assess machine learning (ML) models and long-short term memory model (LSTM) predicting hospital utilisation and health-related quality of life (HRQoL) changes during chemotherapy. The performance of models including and excluding PROMs and developed on data affected by different pre-processing methods were compared. ML models and LSTM built on longitudinal PROMs and symptom severity reports collected weekly were also explored. Consultations with a clinical oncologist and patient representatives were an essential part of the study design.
The results indicate that PROMs can improve the performance of AI models, which encourages the collection and use of patient-reported data in AI research. Using oversampling techniques can affect the performance of models predicting the same outcomes. This thesis recommends approaches which mitigate the risk of bias (including preventing test set from the effects of data pre-processing) and nested cross-validation pipeline for model development and evaluation. HRQoL outcomes were chosen by patients and a clinical oncologist, which emphasises the importance of stakeholder involvement. Finally, the results indicate that LSTM model can detect sudden changes in data to successfully predict hospital utilisation events within 14 days from a completed symptom severity report, but does not outperform simple ML models for long-term predictions.
Based on these results, a patient-centric framework to develop and evaluate clinically relevant AI models applied on static and time-series PROMs is proposed. This framework can help mitigate bias and support the consistent use of valuable PROMs data in AI research predicting patient outcomes.