Modeling trajectories of routine blood tests as dynamic biomarkers for outcome in spinal cord injury

Abstract

Routinely collected blood tests can reflect underlying pathophysiological processes. We demonstrate that the dynamics of routinely collected blood tests hold prediction validity in acute Spinal Cord Injury (SCI). Using MIMIC data (n = 2615) for modeling and TRACK-SCI study data (n = 137) for validation, we identified multiple trajectories for common blood markers. We developed machine learning models for the dynamic prediction of in-hospital mortality, SCI occurrence in spine trauma patients, and SCI severity (motor complete vs. incomplete). The in-hospital mortality model achieved an out-of-train ROC-AUC of 0.79 [0.77-0.81] day one post-injury, improving to 0.89 [0.88-0.89] by day 21. For detecting the presence of SCI after spine trauma, the highest ROC-AUC was 0.71 [0.69-0.72] achieved by day 21. By day seven, the ROC-AUC for SCI severity was 0.81 [0.77-0.85]. Our full models outperformed the severity score SAPS II following seven days of hospitalization.

Fig. 1. Summary of the multi-stage analysis and computational experiments.

a Flowchart of the multi-stage analysis. After cohort selection from MIMIC data using ICD9 and 10 codes, lab values were curated and processed for modeling. The trajectory modeling for each marker is performed by first doing a model search to select the best link function and degrees of freedom for the smoother function (natural cubic spline), followed by a linear search of the number of trajectory classes. The selected final models are used for predicting the probability of trajectory membership per subject at different timepoints from hospitalization. These probabilities are then used as predictors in ML classifiers. b Three prediction experiments were set up to study the potential use of blood trajectories as dynamic biomarkers. c Schematic of the dynamic prediction experiments simulating increased data availability over time after hospital arrival.

Fig. 2. Flow diagram of cohort build.

Patients from MIMIC-III/IV were first filtered based on their ICD9 and ICD10 diagnostic codes. Then, potential overlapping patients were filtered from MIMIC-IV. After laboratory analyte data extraction and data cleaning, patients with less than 3 measures for any of the 20 most common analytes were excluded. The data from a total of 2615 patients from both MIMIC databases were used for trajectory modeling as well as Experiments I and II. An additional 137 patients from the TRACK-SCI dataset were used as a validation set in Experiment III to predict SCI severity based on AIS grade.

Fig. 3. Spaghetti plots for the outlier-cleaned modeling set of laboratory analytes for the first 21 days after admission in the MIMIC data.

Each line represents a single subject. Lines red and blue are two randomly selected subjects illustrating differences in temporal trends.

Fig. 4. Blood trajectories from MIMIC data.

a The mean trajectory for each one the classes for each analyte model are shown, together with the 95% CI. Note that although colored the same for visualization, these are univariate models and therefore classes might not be constituted by the same subjects across analytes. The legend in each panel shows the number of subjects assigned to each trajectory class. b Univariate summary of association of the different trajectory classes with demographics and other variables of interest. The data are shown as a heatmap of the q value (adjusted p value for multiple comparisons). Red boxes indicate a q value < 0.05.

Fig. 5. Out-train ROC-AUC performance of dynamic predictions.

a ROC-AUC out-of-train sample performance of experiment I in-hospital mortality. b ROC-AUC out-of-train sample performance of experiment II for detecting the presence of SCI after spine trauma. c ROC-AUC out-of-train sample performance of experiment III on detecting SCI severity on the TRACK-SCI cohort, external to trajectory modeling. Dashed red lines represent the non-information rate (mean prevalence of the outcome of interest in each experiment). Three predictors’ lists are shown: Traj. PPA = posterior probability of class assignment only; + Sum. stats = addition of summary statistics of blood data; and + BL = addition of baseline predictors.