Short- and long-term weekly patient-reported outcomes prediction undergoing radiotherapy: single-patient time series model vs. transformer-based multi-patient time series model
Abstract Background Patient-reported outcomes (PROs) are direct reports from patients on health status, symptoms, quality of life, or treatment satisfaction, offering critical insights into subjective experiences that clinical metrics may overlook. Accurately predicting personalized short- and long-...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
BMC
2025-08-01
|
| Series: | BioData Mining |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s13040-025-00464-7 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Background Patient-reported outcomes (PROs) are direct reports from patients on health status, symptoms, quality of life, or treatment satisfaction, offering critical insights into subjective experiences that clinical metrics may overlook. Accurately predicting personalized short- and long-term weekly PROs during radiotherapy is essential for monitoring health status, optimizing treatment efficacy, and enabling timely interventions to manage side effects. Methods Based on the well-documented prostate cancer PRO dataset with 17 patients after pre-processing, this study evaluates single-patient time series models (i.e., vector autoregression (VAR) and VAR with incremental ground truth PRO data (VAR-Inc)) and a transformer-based multi-patient model (i.e., Temporal Fusion Transformer (TFT)) for short- and long-term weekly PRO prediction. VAR-Inc integrates follow-up PRO data to refine predictions, while TFT leverages multi-patient heterogeneous information to capture complex temporal patterns. Results Key experimental results on prostate cancer patients demonstrate that (1) VAR-Inc demonstrated superior performance (lower MAE/RMSE) over VAR, highlighting the importance of incremental PRO updates. (2) TFT significantly outperformed both VAR models in long-term prediction, with statistical significance, by utilizing multi-patient data. (3) TFT effectively captured weekly PRO trends and variations, aligning closely with ground truth. (4) Unlike single-patient models, TFT built robust predictive frameworks by integrating cross-patient similarities and complementary patients’ PRO information. VAR-Inc’s performance deteriorated with missing follow-up PROs, whereas TFT remained stable, overcoming this limitation. On average, TFT outperforms VAR and VAR-Inc by achieving a lowest MAE 0.7715, while the MAE of VAR and VAR-Inc are 1.1329 and 0.8089, respectively. Furthermore, TFT is superior to VAR and VAR-Inc by achieving a lowest RMSE 0.9586, while the RMSE of VAR and VAR-Inc are 1.4817 and 1.0693, respectively. Conclusion TFT emerges as a reliable approach for PRO prediction, excelling in long-term accuracy, trend capture, and resilience to data gaps by leveraging multi-patient information. Its ability to synthesize heterogeneous PRO data offers advantages over single-patient models, supporting personalized treatment adaptation and informed clinical decision-making. This underscores the potential of transformer-based models in enhancing PRO-driven radiotherapy management. |
|---|---|
| ISSN: | 1756-0381 |