Leveraging single-sample networks for predictive modelling of longitudinal clinical data

Longitudinal clinical data collected in non-intensive care units (non-ICUs) are characterised by sparsity, irregular sampling, pervasive missingness, and weak temporal structure. Tree ensembles trained on discretised laboratory time series outperform temporal deep learning models in this regime, but treat variables independently. In contrast, clinicians interpret laboratory measurements jointly, evaluating how they evolve over time relative both to one another and to a reference population. We hypothesise that patient-specific shifts in inter-variable relationships provide predictive information beyond individual trajectories and can identify biomarkers whose clinical utility arises from coordinated physiological change rather than isolated values. We test this hypothesis by adapting single-sample networks (SSNs) from network medicine, which quantify how an individual sample perturbs a population-level feature-association structure. Originally developed for static omics data and network module mining, canonical SSNs face additional challenges when applied to sparse, longitudinal clinical measurements in train/test evaluation settings. We address these challenges by extending SSNs along five methodological axes: (i) leakage-free network construction (leave-one-out for training and include-one for held-out evaluation); (ii) configurable edge normalisation and data-adaptive thresholding; (iii) temporal handling through four representations (static, trajectory-summarised, discretised, and stacked); (iv) missing-aware and incremental aggregation via pairwise deletion and Welford's algorithm; and (v) node-, edge-, and graph-level feature extraction for downstream prediction. The framework is modular, allowing new association measures and edge-building functions to be seamlessly integrated. We benchmark the framework across ten datasets: two non-ICU cohorts extracted from MIMIC-IV, two ICU cohorts (PhysioNet 2012 and 2019), and six TCGA omics cohorts. Configurations tailored to longitudinal clinical data (discretised temporal handling, missing-aware aggregation, and edge-wise z-score normalisation without thresholding) substantially outperform canonical SSN defaults, demonstrating the importance of adapting network construction to temporal structure and missingness. Network-derived edge and node features retain 98-99% of the predictive performance of standard representations (discretised time series). Importantly, node-level features exhibit substantially greater feature-importance stability, improving by 19-25% on average across cohorts (longitudinal and static), attribution scopes (fold-level and sample-level), and stability metrics (Kuncheva index and Kendall's W). In a case study predicting chemotherapy-related aplasia and neutropenic fever, node-level features prioritise relatively rare white blood cell subpopulations involved in immune activation and bone-marrow recovery (e.g., eosinophils and myelocytes), suggesting that their predictive value is driven less by their absolute trajectories than by their relationships with the broader haematological profile. Consistent with this interpretation, a size-matched relational (edge) representation of these markers outperforms their raw values (e.g., eosinophils: 0.718 vs. 0.655 ROC-AUC). Overall, relational representations improve feature importance stability and highlight interaction-driven biomarkers overlooked by standard representations while maintaining competitive predictive performance. The framework's extensibility further enables systematic development and evaluation of domain-specific network-construction strategies: two custom edge-building functions that incorporate aggregated and longitudinal missingness patterns yield consistent gains over canonical edge-building functions on the chemotherapy cohorts. These findings support single-sample networks as a flexible framework for predictive modelling and biomarker discovery in longitudinal clinical data. The pipeline is available at: https://anonymous.4open.science/r/longitudinal-patient-networks/.