In this study, we developed and externally validated ML algorithms to predict clinically important impairment in HRQoL within multiple pooled datasets. We report our findings according to the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines [20] and the Guidelines for Developing and Reporting Machine Learning Predictive Models in Biomedical Research [21].

Ethical approval for the secondary data analysis was obtained in 2022 at the Antoni van Leeuwenhoekziekenhuis/Nederlands Kanker Instituut from the institutional review board (IRBd22-179).

For the cohorts: ethical approval for the OPTIMUM (Towards Optimal Timing and Method for Promoting Sustained Adherence to Lifestyle and Body Weight Recommendations in Postmenopausal Breast Cancer Survivors) study was obtained from the Medical Research Ethics Committee Brabant, the Netherlands (NL66913.028.18). The VERDI (Verlauf der Diagnostischen Abklärung) study was approved by the ethics committees of the University of Heidelberg and the Medical Association of Saarland, Germany. Written informed consent was obtained from all participants. The study protocol for UMBRELLA (Utrecht Cohort for Multiple Breast Cancer Intervention Studies and Long-Term Evaluation) was approved by the Institutional Review and Ethics Board of the University Medical Center Utrecht, the Netherlands. No further compensation for the original participants was provided as their identity was not known to us. Patient data was processed only in pseudonymized form.

We described the data pooling previously in more detail [11]. In brief, the BALANCE dataset comprises 6 cohorts with a total of 6316 female patients with nonmetastatic breast cancer, including trial data (EORTC [European Organisation for Research and Treatment of Cancer] AMAROS [After Mapping of the Axilla, Radiotherapy or Surgery?] [22]), RWD (Netherlands Cancer Institute [NKI] [23], district hospital Kufstein, Austria [24]), and observational studies (UMBRELLA [25], OPTIMUM [26], and VERDI [27]), with at least 2 HRQoL assessments (Figure 1). Data were collected between 2001 and 2024 and encompass patients receiving active treatment, follow-up, or survivorship care.

Data cleaning and preprocessing were conducted using R (R Foundation) [28], Jupyter Notebook, and Python (Python Software Foundation) libraries (Methods S1) in Multimedia Appendix 1 were used for statistical analyses and ML algorithms.

Features include patient characteristics and sociodemographics, comorbidities, disease and treatment parameters, as well as the items of the EORTC QLQ (Quality of Life Core Questionnaire)-C30 [29] and breast cancer module QLQ-BR23 [30]. Variables were either measured at the time of diagnosis or at the first HRQoL assessment. Cohort data were harmonized based on the common codebook (see Methods S2) in Multimedia Appendix 1. Briefly, time-related variables were coded in days since the date of diagnosis to ensure consistency across cohorts. Treatment characteristics were dichotomized into binary categories, and comorbidities were summarized as simple counts derived from available diagnostic information. Disease-related variables (eg, tumor stage and receptor status) were largely consistent across cohorts.

Features were preprocessed by one-hot encoding categorical variables (missing values were treated as distinct categories) to ensure they are represented numerically without implying an ordinal relationship and by normalizing all numeric features to 0 mean and unit variance to standardize their range, improving model performance and comparability. We removed features with 0 or near-0 variance (<0.001) from the analysis and imputed missing feature values using the iterative imputer implementation from scikit-learn. It estimates missing values in an iterated round-robin fashion as a function of other features based on a Bayesian ridge estimator [31].

To evaluate time-dynamic (ie, predicting HRQoL at every available time point using prior data and time since last assessment) predictive performance, for each patient, all possible combinations of their assessments were compiled into distinct pairs. Observations lacking either HRQoL data or the corresponding time point information were excluded. A new variable was introduced to represent the difference in days between the assessment dates within each pair. The outcome variable for each pair was defined as the value from the later assessment. For example (see Methods S3 in Multimedia Appendix 1 for an illustration), if a patient had 3 assessments, A1, A2, and A3, occurring at different time points, pairs such as (A1, A2), (A1, A3), and (A2, A3) were generated. The earlier assessment in each pair contributed its single items to the dataset, while the later assessment provided the target outcome based on the dichotomized scales (see subsection Outcomes under the Methods section). No other data from the second assessment in each pair was included in the input features. Each pair (A_m, A_n) is represented as a training sample (X_m, t_m-n, Y_n), where X_m are the features of the assessment m, Y_n is the binary outcome variable of the respective observation, and t_m-n is the time difference between A_m and A_n in days. Each training sample can be interpreted as “t_m-n days after the patient has reported results X_m, the patient’s outcome is Y_n.” This approach allowed us to model changes in HRQoL over time. Figure S1 in Multimedia Appendix 1 shows the distribution of the time difference between assessment pairs in the train and test set.

HRQoL was measured with the EORTC QLQ-C30 before, during, or after treatment. Symptom (fatigue, nausea and vomiting, pain, dyspnea, insomnia, appetite loss [AP], constipation, diarrhea, and financial difficulties) and functioning scales (physical, role, emotional, cognitive, and social) are reported on a 4-point Likert scale (“not at all,” “a bit,” “quite a bit,” and “very much”) and transformed to linear scores ranging from 0 to 100 scale [32]. We dichotomized the outcomes according to the established thresholds for clinical importance to indicate clinically relevant impairments, defined as scores associated with at least one of the following patient-reported concerns: limitations in daily life, need for help or care, and worries of the patient or their partner or family [33]. The dichotomization simplifies clinical interpretation through established and meaningful thresholds.

We did not adjust the dataset for class imbalance as most outcome variables were balanced (see Results section), and adjusting for class imbalance would increase the risk for overfitting.

We divided our dataset into model development data (5 cohorts) and external validation data (NKI). The NKI set was chosen as it contains the most current data from clinical practice and a sufficient number of patients.

We used the scikit-learn implementations of algorithms from multiple classifier families [34] that proved effective in similar studies [6,12,35,36]. We evaluated a logistic regression with L2 regularization, an extra-trees classifier, a multilayer perceptron classifier, and a histogram-based gradient boosting classification tree (see Methods S3 in Multimedia Appendix 1 for a detailed description of libraries and GitHub for the code base). To account for the unequal contribution of patients with varying numbers of assessment pairs, we additionally incorporated sample weights inversely proportional to the number of pairs per patient. The final model was recalibrated using a probability calibration with logistic regression from scikit-learn (CalibratedClassifierCV, default settings) using a 5-fold cross-validation based on the training data. In addition, we performed post hoc recalibration on the external validation set to illustrate the extent to which context-specific adjustment can enhance model performance before deployment in a clinical setting.

To assess model performance, we used internal cross-validation using a 5-fold (K=5) strategy. Given the differences in sample sizes across cohorts, the internal cross-validation was configured to keep the distribution of the cohorts within the train and test splits consistent. Thereby, the internal cross-validation was performed in a way that ensures an 80/20 train and test split and also prevents any patient overlap between train and test data.

Model performance was primarily evaluated using the area under the receiver operating characteristic curve (AUC), which is robust against class imbalance [37]. Training with AUC can provide acceptable performance at a 1/10 imbalance ratio without rebalancing [38]. Additional metrics included weighted F₁-score, accuracy, balanced accuracy, and confusion matrix metrics (sensitivity, specificity, positive predictive value, and negative predictive value) based on their scikit-learn implementation using the mean and SD from all internal cross-validation iterations [34]. Accuracy reflects the overall proportion of correct predictions but can be misleading when classes are imbalanced. Balanced accuracy addresses this by averaging the recall obtained on each class, giving equal weight regardless of class frequency. The F₁-score summarizes overall classification performance by balancing false positives and false negatives, with the weighted F₁-score giving more weight to classes with more instances. Sensitivity (recall) quantifies the model’s ability to correctly identify patients with an impairment, whereas specificity reflects the ability to correctly identify those without impairment. Positive predictive value indicates the probability that patients predicted to have an impairment truly experience it, while negative predictive value reflects the probability that patients predicted not to have an impairment are indeed unaffected. Higher scores indicate better overall model performance across all metrics [39].

Last observation carried forward (LOCF) was used as a baseline for model comparison. LOCF serves as a straightforward method for estimating future values where patients’ HRQoL is not expected to change. These absolute values were used to compute AUC scores. It has been used in clinical trials and longitudinal studies as an imputation method or to predict future HRQoL [40,41]. Using this simple approach, therefore, provides a better baseline performance threshold compared to random chance (AUC=0.5). Furthermore, simplified logistic regression models with the prior HRQoL and the time difference as their only 2 predictors were computed to illustrate gains beyond LOCF.

Based on the internal cross-validation, we used the best-performing model for each target variable for external validation in the NKI dataset (Figure 1). Bootstrapping over 1000 iterations was used to estimate the stability of the external validation and LOCF, including 95% CIs [42].

To interpret model predictions, we applied permutation feature importance, quantifying the contribution of each feature to the trained models. Model calibration was evaluated with calibration plots, the calibration plot’s slope and intercept, and the expected calibration error (ECE), which assesses how closely predicted probabilities align with observed outcomes; good calibration is indicated by a calibration curve close to the diagonal, an intercept near zero, a slope approaching one, and a low ECE [39,43]. The ECE was based on a kernel-smoothed function using the relplot Python library [44]. Additionally, the Brier score was computed to assess the overall performance, combining discrimination and calibration, with lower values near 0 representing better performance [45].

Decision curve analyses were conducted to determine the clinical utility of the models [46]. Predicted risks may inform patient counseling, leading to heterogeneous downstream actions, such as additional assessment, supportive care, or monitoring, which differ in burden and resource use and cannot be directly compared. Evaluating net benefit across a wide threshold range (ie, the predicted risk at which a clinician would consider acting), therefore, allows assessment of model utility under varying clinical preferences and use scenarios. In this framework, the treat-all and treat-none strategies represent reference scenarios in which all patients or no patients, respectively, are considered at risk. A model is considered to have clinical utility at threshold probabilities where its decision curve yields a higher net benefit than both reference strategies, indicating more favorable trade-offs between identifying patients at risk and the use of health care resources [47].

In line with our aim of developing fair HRQoL prediction tools, and in accordance with recommendations to ensure equitable model performance across diverse populations [16], we defined risk groups within the external validation dataset to address group fairness concerns. The following risk groups were selected based on prior literature [48] and evaluated in the same manner as the full external dataset: (1) postmenopause, (2) financial difficulties (according to the thresholds for clinical importance for the financial difficulties scale [33] at baseline), (3) obesity (BMI≥30), (4) 2 or more comorbidities, (5) lower educational status (secondary education or lower), and (6) frailty (scoring according to Murugappan et al [49]). Additionally, to the previously described metrics, we evaluated true and false positive rates to screen for equalized odds disparities.

Further, we split the validation dataset to assess the impact of the time-dynamic aspect on model performance. As we are not relying on fixed prediction intervals, we want to ensure stable model performance across different time horizons. We evaluated long-term prediction defined as predicting outcomes more than 1 year in the future (t_m-n larger than 365 days), as such predictions are especially valuable for informing survivorship care planning and long-term patient management. Further, we selected all outcome pairs from within the first year after diagnosis, as this period typically encompasses active treatment, during which most changes in HRQoL are expected to occur. After the first year, HRQoL tends to stabilize, making early assessment particularly relevant for capturing clinically meaningful variation [50,51].

Outcome data from 6316 patients were initially acquired. After cleaning and preprocessing, 4491 patients remained in the model development dataset (Figure 1). Table 1 shows their main characteristics across all included datasets.

The outcome distribution in the development and validation sets after pairing HRQoL assessments showed a higher prevalence of impairments to HRQoL in the external validation set across all domains compared to the training set (Table 2). Major class imbalance (ratio of event to total number of less than 1/10) occurred in 3/14 scales (AP, constipation, and social functioning) in the training set. The event rates were higher in the external validation set.

Figure 2A illustrates the performance of the logistic regression, extra-trees classifier, multilayer perceptron classifier, and a histogram-based gradient boosting classification tree for the prediction of impaired HRQoL at the next available assessment. Across all evaluated QLQ-C30 domains, the histogram-based gradient boosting classification tree showed the highest AUC, and all algorithms outperformed the LOCF baseline (Figure 2B). AUC values for the baseline regression models lay between 0.648 (95% CI 0.616‐0.679; pain) and 0.829 (95% CI 0.767‐0.884; AP), see Figure 2B for a comparison. A comprehensive list of evaluation metrics can be found in Table S1 in Multimedia Appendix 1.

From all QLQ-C30 domains, functional scales had a higher AUC than single-item symptom scales (dyspnea, insomnia, AP, constipation, and diarrhea). Except for financial difficulties, which showed the highest AUC overall (0.861, SD 0.009).

The external validation with the NKI yielded good discriminatory ability (Figure 2B and Table S2 in Multimedia Appendix 1). The highest AUC (0.837, 95% CI 0.809, 0.864) could be observed in the financial difficulties scale, whereas the diarrhea scale yielded the lowest AUC (0.643, 95% CI 0.600, 0.689).

Model calibration varied between the scales (Figure S2) in Multimedia Appendix 1. We observed the worst calibration in the single-item symptom scales (AP, constipation, and diarrhea), better calibration in the functioning scales, and the best calibration in fatigue and physical functioning. Generally, models for scales containing more items were better calibrated. Calibration slopes were below the ideal value of 1 for all models (Table S2) in Multimedia Appendix 1, ranging from 0.089 (constipation; 95% CI 0.063, 0.116) to 0.286 (fatigue; 95% CI 0.259, 0.313), indicating overconfident predictions. Post hoc recalibration improved discrimination (AUCs 0.975‐0.995) as well as calibration based on visual inspection (Table S5 and Figure S3) in Multimedia Appendix 1.

Across all domain models, the individual HRQoL items from the previous assessment, the time of assessment, and treatment start dates were among the most important features (Figure S4) in Multimedia Appendix 1. The type of treatment, other clinical variables, and diagnostic characteristics or comorbidities had little influence on the models’ performance.

Decision curves are shown in Figure S5 in Multimedia Appendix 1, and most models (exceptions: social functioning, AP, and diarrhea) proved to be superior compared to a “treat all” scenario for risk thresholds between 0.2 and 0.6.

The proportion of events in the risk groups was almost always higher than in the external validation set (Table S3) in Multimedia Appendix 1. Figure 3A illustrates model performance in risk groups selected from the external validation set in comparison with the whole set (see Table S3 in Multimedia Appendix 1 for risk group distributions).

The discrimination power within risk groups varied between domains, depending on specific risk group-scale combinations (Table S4) in Multimedia Appendix 1. The lower education status risk group showed the highest (financial difficulties: 0.897, 95% CI 0.825, 0.956) as well as the lowest (nausea and vomiting: 0.473, 95% CI 0.314, 0.633) AUC over all.

In the postmenopausal and obese risk groups, AUC CIs consistently overlapped with those of the full dataset, indicating similar model performance. This was largely consistent for other risk groups, except for the following cases: multiple comorbidities (pain and dyspnea), financial difficulties, frailty (AP and role functioning), and lower educational status (nausea and vomiting, and constipation), where reduced performance was observed.

Additionally, the predictive performance differed within some scales more than in others. For instance, in the pain scale, the AUCs were relatively homogenous, whereas the constipation scale had a high variability in AUCs. Model calibration was poor with calibration slopes across all targets and risk groups below the ideal value of 1 (Table S4) in Multimedia Appendix 1. Due to the low number of events, we could not evaluate the calibration plots.

We observed little impact of the time-dynamic aspect on model performance (Figure 3B). The prediction within the first year after diagnosis showed similar AUCs compared to the whole external dataset, with differences below 0.01 (Table S4) in Multimedia Appendix 1. Long-term prediction showed slightly higher AUCs in 9 of 14 cases and the greatest deterioration in role functioning (0.032).

This study developed the first time-dynamic machine ML model for predicting impaired HRQoL in patients with early breast cancer, using multiple cohorts from different countries and settings. The ML-based prediction models consistently outperformed LOCF and the regression baseline in 10/14 cases, with the histogram-based gradient boosting classification tree showing the best performance. The best calibration was seen in multi-item scales; however, predictions were overconfident in all domains. We introduce a novel approach to time-dynamic HRQoL prediction, enabling flexible predictions beyond fixed time points. This addresses the issue with current static models, as they cannot capture the variability of real-world HRQoL assessment schedules. It therefore allows the early detection of meaningful HRQoL impairments throughout a patient’s course of disease and provides clinicians with a tool to guide targeted interventions. Model fairness analyses revealed variation in model performance across risk groups, depending on individual domain and group combinations.

Overall, our models demonstrated good discriminatory performance across most HRQoL domains; however, calibration was inconsistent, particularly for scales with low outcome prevalence. Discrimination, reflected in AUC values exceeding 0.75 [52], was generally strong, confirming that the models effectively distinguished between patients with impaired HRQoL and those without. This is within the range of comparable studies, most of which used internal k-fold cross-validation (see Krepper et al [12] for an overview). Yet, for some outcomes (especially AP, constipation, and diarrhea) predicted probabilities did not consistently align with observed event rates (ie, poor calibration), limiting the clinical reliability of the models in those domains.

Accurate predictions, based on both strong discrimination and acceptable calibration, were observed for key domains such as physical, emotional, and cognitive functioning, as well as fatigue, pain, and financial difficulties. These domains might reflect more persistent problems, likely contributing to the robustness of the predictive models [53,54]. In contrast, models struggled to predict transient or low-prevalence symptoms, including diarrhea, nausea and vomiting, and constipation. These limitations reflect both the nature of the symptoms, often short-lived and influenced by acute factors [55], and the statistical challenges of modeling rare outcomes [39]. The poor calibration in these domains can be attributed to the class imbalance and the insufficient total number of events. In our external dataset, some of these outcomes (AP and constipation) occurred fewer than 200 times, a known threshold below which reliable calibration curves become difficult to estimate [43], while also corresponding to less than 10% of all observations. Here, ensemble methods such as random forests and gradient boosting tend to struggle, often underestimating probabilities at the extremes due to their inherent smoothing behavior [56]. This happens because averaging predictions from many noisy base models pulls extreme values toward the center, making it difficult to produce confident predictions near 0 or 1.

To address these limitations, we recalibrated the models using data from the specific clinical setting where they could be applied, thereby adjusting for local outcome prevalence and population characteristics. We highlighted the impact of this approach in a supplementary analysis, which improved model performance drastically.

Another potential approach is upsampling the training data to balance class distributions, which showed improved calibration when using HRQoL to predict clinical events [36]. However, this does not reflect real-world incidence rates and may distort probability calibration without improving model performance in cases with an adequate number of events. Notably, similar challenges have been reported in other HRQoL modeling efforts: for example, Adiprakoso et al [35] observed good discrimination but suboptimal calibration when predicting fatigue in patients with cancer, reinforcing the need for careful evaluation of both performance dimensions.

Despite these issues, this study adds to the growing body of literature emphasizing the value of external validation, a step still rarely undertaken in HRQoL modeling despite TRIPOD guidelines recommending distinct datasets for model evaluation [20].

In line with previous research [35,57-59], prior HRQoL was consistently among the most important predictors across all models, even over longer periods of time. Temporal variables, such as treatment start and stop dates and the timing of HRQoL assessment, also ranked highly. This was also reflected in the competitive performance of the baseline model, which included only prior HRQoL and the time between assessments. In contrast, neither specific treatment types nor clinical variables (eg, tumor stage) appeared influential for predicting the general functioning and symptom domains of the QLQ-C30. This suggests that, for broad HRQoL outcomes, being on or off treatment may matter more than which specific treatment is received. These patterns may differ for treatment-specific side effects not captured by the QLQ-C30 (eg, skin problems for new targeted agents), highlighting an avenue for future studies using more granular or treatment-targeted patient-reported outcome (PRO) measures [60].

Decision curve analysis indicated that most models provided greater net benefit than treat-all and treat-none strategies across threshold probabilities between 0.2 and 0.6, suggesting potential clinical usefulness within this range. As subsequent actions following risk identification are not clearly defined and may vary widely (see Di Meglio et al [57] for a list of suggested interventions for patients with high risk of fatigue), the models should not be interpreted as directly guiding specific interventions. Rather, given that available options are often low risk and primarily involve the use of supportive resources, the models’ main value may lie in informing patient counseling and facilitating early, shared discussions about potential support needs.

Model fairness analyses revealed disparities in model performance, notably among patients with lower educational status or frailty. However, there was no single risk group with the constant worst discrimination in all domains. Our findings echo concerns about algorithmic bias in health care ML models, exacerbating existing inequities in socioeconomic status, disability, and other characteristics [61-63]. This was reflected in the increased prevalence of impaired HRQoL across the selected risk groups, which nevertheless could not always be predicted correctly. Addressing these risks requires the use of fairness-aware ML techniques, such as reweighting, adversarial debiasing, or subgroup-specific calibration, to ensure equitable performance across diverse patient populations [17]. Future work should explore these strategies to prevent the amplification of health care disparities.

Our implementation of time-dynamic prediction marks a promising methodological extension over traditional fixed-interval ML models for HRQoL. Enabled by a heterogeneous dataset, this approach better reflects real-world clinical care, where follow-up intervals vary. Importantly, we observed consistent model performance across different time spans. For long-term prediction, that is, spanning more than 1 year between assessments, this might reflect a stabilization of HRQoL. At the same time, the models also performed well during the first year after diagnosis, despite this period being marked by significant treatment-related fluctuations in HRQoL. This indicates that the models are capable of capturing both stable and dynamic phases of a patient’s trajectory, making them applicable in both contexts. In line with the ESMO (European Society for Medical Oncology) clinical practice guideline for PRO use in clinical practice, which recommends integrating PRO results into clinical care [64], such insights may inform supportive care planning by enabling the prioritization of patients who are likely to require timely interventions, while also helping patients develop more realistic expectations regarding their long-term HRQoL prognosis.

As our approach primarily served as a proof-of-concept, no direct comparisons with existing HRQoL models are currently available. Nevertheless, these encouraging results highlight the potential of time-dynamic modeling to address real-world clinical needs, warranting extension to other populations and settings. Future research is needed to compare this approach to static models.

This study presents several key strengths. To our knowledge, it is the first ML approach to HRQoL prediction that incorporates time-dynamic modeling for patients with breast cancer, allowing for predictions at variable future time points rather than fixed intervals. Second, deriving our outcomes directly from anchor-based thresholds (patients reporting limitations in daily life, need for help or care, or worries of the patient or their partner or family) [33] ensures that predictions are tied to clinically meaningful impairments. As such, the models can help identify which patients are more likely to develop problems in specific HRQoL domains at selected points in the future and who may benefit from further clinical attention. Considering that PRO data are increasingly collected in routine care and that the QLQ-C30 is among the most frequently used questionnaires in electronic systems [65], our study shows that these data could be used more comprehensively to inform clinical teams and patients. Additionally, the use of a multicohort dataset, combining clinical trial data with real-world observational data, improves the model’s validity. Importantly, we used external validation using a separate cohort to evaluate generalizability and illustrated the effects of post hoc recalibration. We also explicitly assessed predictive fairness, addressing a critical and often overlooked aspect of ML in health care by examining performance across subgroups and discussing bias mitigation strategies.

Nevertheless, this study has limitations. Most patients in our dataset were Dutch, which may limit the generalizability of findings to other countries. However, reference values for early patients with breast cancer demonstrate only slight variations across Europe [66], and the models primarily rely on prior HRQoL rather than population-specific variables. This might suggest that their performance is not inherently tied to national context, and the models may be transferable to other similar countries, provided that health care systems, cultural backgrounds, and follow-up intensity are broadly comparable. Yet, model adoption into different settings must still be evaluated carefully. In addition, we are currently working to include further German and British data to expand our database and retrain our models with a more international population.

Second, even though we evaluated model performance in an RWD set and this might have reduced selection bias, there might still be an underrepresentation of certain minorities or risk groups in our data.

Furthermore, as we performed a secondary analysis on existing datasets not originally intended for this purpose, some potentially important predictors of HRQoL (eg, socioeconomic status or social support) were unavailable, which may have limited the performance of our models.

Additionally, we dichotomized the outcomes to facilitate clinical interpretation; this may have led to a loss of information and reduced predictive accuracy compared to modeling with the original continuous scale.

Further, the field of ML is evolving rapidly, and by the time results are published, newer architectures (eg, pretrained tabular foundation models [67]) may already offer improved performance. Exploring such approaches was beyond the scope of this study but represents an important direction for future research.

Finally, while our models demonstrate promise, they remain preliminary with respect to clinical implementation, especially for a few selected scales showing poor calibration for certain rare symptoms. This limitation may be attributable not only to low event rates but also to the application of a single time-dynamic modeling framework across heterogeneous phases of the patient journey, raising the possibility that more phase-specific approaches (eg, during active treatment vs posttreatment or survivorship) could improve calibration for certain symptoms.

This mirrors a systemic issue in translational implementation, where HRQoL prediction models often fail to progress beyond proof-of-concept due to a combination of technical, operational, and contextual barriers [68]. As outlined by Spencer et al [69], both general ML deployment issues and PRO-specific challenges contribute to what they refer to as the “leaky pipeline” from model development to bedside application. Moreover, compliance with regulatory frameworks, such as the European Union Artificial Intelligence Act, necessitates not only robust prospective validation in real-world settings but also demonstrable alignment with ethical, safety, and stakeholder-involvement standards.

Our findings support the future integration of ML into PRO-based clinical decision support systems, while also emphasizing the need to address fairness in predictive modeling. Importantly, the predicted impairments are clinically meaningful, as they identify patients at risk of future problems derived from anchor-based thresholds and enable more timely, targeted interventions. The time-dynamic nature of the predictions further increases their clinical utility by allowing risk estimation to be aligned with clinically relevant time horizons rather than fixed follow-up intervals. Future work should focus on improving model fairness, expanding dataset diversity, and evaluating clinical implementation in real-world oncology settings.