Unsupervised machine learning models reveal predictive markers of glioblastoma patient survival using white blood cell counts prior to initiating chemoradiation

Purpose: Glioblastoma is a malignant brain tumor requiring careful clinical monitoring even after primary management. Personalized medicine has suggested use of various molecular biomarkers as predictors of patient prognosis or factors utilized for clinical decision making. However, the accessibility of such molecular testing poses a constraint for various institutes requiring identification of low-cost predictive biomarkers to ensure equitable care. Methods: We collected retrospective data from patients seen at Ohio State University, University of Mississippi, Barretos Cancer Hospital (Brazil), and FLENI (Argentina) who were managed for glioblastoma—amounting to nearly 600 patient records documented using REDCap. Patients were evaluated using an unsupervised machine learning approach comprised of dimensionality reduction and eigenvector analysis to visualize the inter-relationship of collected clinical features. Results: We discovered that white blood cell count of a patient during baseline planning for treatment was predictive of overall survival with an over 6-month median survival difference between the upper and lower quartiles of white blood cell count. By utilizing an objective PDL-1 immunohistochemistry quantification algorithm, we were further able to identify an increase in PDL-1 expression in glioblastoma patients with high white blood cell counts. Conclusion: These findings suggest that in a subset of glioblastoma patients the incorporation of white blood cell count and PDL-1 expression in the brain tumor biopsy as simple biomarkers predicting glioblastoma patient survival. Moreover, use of machine learning models allows us to visualize complex clinical datasets to uncover novel clinical relationships.


Introduction
Glioblastoma (GB) patients suffer from aggressive solid tumors of the central nervous system, with 95% projected to be deceased within 5 years following diagnosis (1). Cause of death in glioblastoma varies, but may include herniation secondary to mass effect, treatment complications, and aspiration pneumonia due to brainstem dysfunction (2). Despite aggressive adjuvant therapy, however, almost all patients will experience tumor progression. Nevertheless, major strides have been made within neurooncology to merge personalized medicine to better predict patient outcomes. Although therapies for GB management after maximal-safe surgical resection have remained largely unchanged since the concomitant use of temozolomide (TMZ) with radiation (ChemoRT), molecular biomarkers have been integrated as critical tools in the diagnosis and prognostication of gliomas (3)(4). From a diagnostic standpoint, lack of an IDH-mutation de nes the current World Health Organization (WHO) 2021 de nition of a GB (4). Moreover, TERT promoter mutations, EGFR ampli cations, and modi cations of chromosome 7/9/10 are all recognized molecular changes present in GB.
Prognostically, several molecular markers have been suggestive of predicting outcome in GB patients.
Namely, MGMT promoter methylation status is routinely tested to predict e cacy of temozolomide treatment due to the antagonistic role of MGMT in temozolomide DNA alkylation (5)(6). However, other markers such as CDKN2A/B loss and EGFRvIII have been shown in retrospective studies to act as poor prognostic markers in GB patients or subset populations (7)(8). The need for prognostic markers of disease is critical in ascertaining whether certain patients should be monitored more closely during follow-up. Furthermore, the logistics of e ciently scheduling patients between visits with neuro-oncology, neurosurgery, neuroradiology, and radiation oncology would signi cantly bene t from improved triaging methods that prioritize the most sensitive patients. Unfortunately, the accessibility of molecular markers is not equal across healthcare ecosystems (9)(10). Current discussions in the eld of neuro-oncology have pointed to major accessibility barriers and bioethical implications of pure reliance of molecular biomarkers of disease to understand cancer (10). Molecular pathology testing requires batching to reduce patient costs, a need that results in signi cant costs in turn-around-time for molecular assays (11). In consequence, groups have sought ways to evaluate outcomes of glioma patients using surrogate measures such as neuro-cognitive testing, psychiatric examination, or image analysis of histology to predict molecular phenotypes as these turn-around-times are superior to those of molecular pathology (9,(12)(13). While promising however, integration of these approaches will take time whereas routinely collected data is readily available.
To better explore potential prognostic markers already routinely collected while managing GB patients, we retrospectively evaluated 581 patients at four sites from three countries: Ohio State University (US), University of Mississippi (US), Barretos Cancer Hospital (Brazil), and FLENI (Argentina). Patient features were assessed using an unsupervised learning approach to understand how clinical metrics related amongst each other-including relevant clinical endpoints such as progression-free survival (PFS) and overall survival (OS). Furthermore, by utilizing these novel relationships, we better characterized the relevance of routine complete blood counts (CBCs) and merged its utility with routine IHC staining in pathology to suggest novel work ows which may predict outcomes of patients with GB. approval. Clinical records from patients receiving GB care from 2012 to 2020 were evaluated. Inclusion criteria for patients were designated by prior history of GB treatment. We however, excluded patients with prior history of low-grade gliomas, IDH-mutation, or patients with external care prior or following contact which led to poor documentation of disease course. Study data was collected and managed using REDCap electronic data capture tools (14)(15). REDCap was utilized to (a) securely store and deidentify patient records for downstream use, (b) ensure consistency of data collection, and (c) be distributable to collaborators desiring to replicate or collect similar metrics as described in the study (Supplementary Document 1).

Selection Criteria and Clinical Data Collection
Clinical features were collected as listed from the electronic health record. Comorbidity scoring was performed by manually listing known comorbidities and scoring them in REDCap following the Charlson Comorbidity Index (CCI) (16). Lesion and molecular features were extracted from radiology and pathology reports, respectively. Results from complete blood counts (CBCs)-including white blood cell count (WBC), neutrophil count, lymphocyte count, and platelet count were designated as CBC draws occurring approximately 2-4 weeks after surgery during patient follow-up with neuro-oncology prior to beginning ChemoRT. Neutrophil-to-lymphocyte ratio (NLR) was calculated by dividing the neutrophil count against lymphocyte count in a patient. Steroid dose was de ned as the total daily steroid intake during the same day as CBC collection. Variables containing dates were converted to deidenti ed values of time to remove potential patient timeline identi cation after calculating relevant timespans in days. OS in the study was de ned as the time from primary tumor resection until time of death. PFS was de ned as time from primary tumor resection until time of initial detection of a novel enhancing lesion on imaging. Con rmation of enhancement as being either cancer recurrent or treatment reactive in nature was completed by clinical correlation and consensus from a multi-disciplinary tumor board.

Exploration of Clinical Features
Collected clinical data from REDCap was exported and deidenti ed for use in R. Missingness of data was evaluated using the naniar package (17). Cases with over 30% missing records were not evaluated using unsupervised analysis while any missing features in the remainder of cases were imputed using the mice package in R (Supplementary Fig. 1; 18). Using the factoextra package, principal component analysis (PCA) was performed over the dataset and relations of clinical features were visualized using eigenvector plotting of our PCA as described in our previous work (19)(20).
Exploration of clinical features as a function of OS was explored using cox regression modeling with nonimputed data. Clinical features were independently evaluated using a univariate cox model to assess prediction of overall survival time. P-values less then 0.05 were omitted in subsequent modeling. Univariate-signi cant features were then evaluated using a multivariate cox regression model using patient records found to have no missing relevant features. Patient survival was further explored by quartile strati cation of patients for relevant clinical features (Supplementary Table 1). Highest and lowest 25% groups were compared for progression-free survival and overall survival using the survival and survminer packages (21)(22). Visualization of relevant clinical confounders was performed using R with ggplot2 (23).

PDL-1 Image Analysis
PDL-1 immunohistochemistry imaging was collected from relevant patient samples at Ohio State which underwent routine PDL-1 evaluation at time of primary surgery for GB. Image tiles were collected from digital pathology slides and processed in R using EBImage (24). RBG images were deconvoluted for hematoxylin and DAB stain layers using scikit-image and reticulate (25-26). Segmentation for either hematoxylin or DAB staining was performed using Otsu thresholding (27). Segmented regions were calculated for morphologic features using EBImage. Filtration of segmentation was performed using a random-forest based classi er trained over intensity-based morphology features to classify segmentation as no stain, low stain, medium stain, and high stain (Supplementary Fig. 2; 28). No stain segmentations were ltered. PDL-1 staining was represented by the ratio of DAB-stained pixel to hematoxylin-stained pixels-calculated from segmentation areas in EBImage based upon objective stain scoring approaches delineated by Igarashi et al (29).

Patient Characteristics
A total of 581 patients were evaluated and documented in REDCap across four centers. In turn, overall demographic scores and biases across centers were assessed. Overall, mean age was 61 years [range:   (Table 1). Males slightly exceeded females in representation across centers (Fig. 1A). The distribution of ethnic origin directly related to the site of collection. North American sites represented predominately Caucasians and Black/African Americans with non-Latino ethnicity, while South American sites were of Caucasian and mixed ancestry with Latino origin (Fig. 1B-C). All patients had biopsy con rmed GB with resection of tumor at primary surgery. There was a slight predominance of left sided lesions primarily occurring in the frontal, temporal, and parietal lobes across sites ( Fig. 1D-E). On average, mean tumor diameter was 4.34 cm [range: 0.1-9.5] with most without midline shift on imaging (Table 1).

Unsupervised Analysis of Patient Reveals Capabilities of CBCs in Predicting Overall Survival Outcome and Time to Enhancement
Utilizing the various collected clinical datapoints, we sought to evaluate whether speci c clinical features were correlated with relevant outcomes in patients that have not been previously integrated in clinical practice. Speci cally, we posited that applying an unsupervised machine learning approach could represent clinical relationships amongst features to allow us to visualize novel ndings predictive in GB patient prognosis. To do so, our multidimensional dataset was reduced using principal components analysis (PCA) to stratify patients by these clinical metrics and visualized using PCA eigenvector plotting (Fig. 3) . 4A-B). Furthermore, evaluation of OS using WBC load and neutrophils showed similar trends between groups (p = 0.00042 and p = 0.0007, respectively). However, while NLR did show signi cant survival difference in OS (p = 0.0081), the difference in PFS between groups did not reach our threshold of signi cance (Fig. 4C).
Combined with the previous analyses, our ndings strongly support the use of CBCs in predicting both OS and PFS in patients. Speci cally, the evaluation of routinely drawn WBC and neutrophil count at time of pre-ChemoRT planning is shown to be correlative with poorer survival outcomes in patients with high load compared to patients with low load.

WBC Load is Re ective of Intrinsic Tumor Microenvironment Changes Present in Glioblastoma
As the measures of WBC load evaluate circulating immune counts, we posited these differences in peripheral immune activity may correlate with intrinsic tumor microenvironment differences found in primary GB events. To assess this, we selected PDL-1 IHC staining done in a subset of patient during clinical evaluation. Amongst our study population, 57 cases had been evaluated by neuropathology for PDL-1 and representative images were collected from cases and segmented using computer vision to quantify staining (Fig. 5A). Due to the membranous staining of PDL-1 however, staining was quanti ed as the ratio of DAB-positive PDL-1 staining against hematoxylin nuclear staining to control for cellularity of tissue. In turn, increased detection of PDL-1 staining is represented by an increased ratio. With this approach, we predicted the detection of PDL-1 staining ratio to be elevated in the WBC-hi group due to the associated poor prognostic outcome of increased PDL-1 staining. In fact, it was observed that the WBC-hi group showed higher ratios of PDL-1 DAB to hematoxylin staining when compared to WBC-lo (p = 0.027, Fig. 5B). Furthermore, assessing the two sides of the ratio comparison it was seen that while the amount of DAB pixels detected in an image was higher in the WBC-hi group (p = 0.037), the detection of hematoxylin pixels did not vary between groups (p = 0.63) (Fig. 5C-D). In conclusion, it could be inferred the increased detection of PDL-1 staining ratio in the WBC-hi group was not a product of increased cellularity as the distribution of hematoxylin was not different. Overall, these ndings infer that an increase in PDL-1 staining was correlated with the WBC-hi group which showed poorer survival outcomes in patients-validating that the observed differences in WBC load are correlative to initial immune activity present in GB lesions at resection.

Steroid Tapering is Highly Heterogenous Following Surgery and May In uence WBC Load
Although our analyses found strong correlation of CBC load to survival outcomes, we further evaluated potential clinical confounders which may in uence CBC levels prior to ChemoRT. As shown in the analyses of our study population, heterogeneity in patient demographics, lesion characteristics, and patient management was present (Table 1). In turn, an assessment to identify whether speci c clinical features signi cantly varied between our hi and lo populations was critical. As patient age, CCI score, and MGMT methylation status were predictive of survival in our univariate cox model, we assessed these factors in addition to other clinical features that were shown to be correlative to CBC measures in the eigenvector plot [lesion size, steroid intake, and BMI] (Fig. 3& Table 2). Comparing WBC-hi and lo groups, MGMT methylation status distribution was not found to signi cantly vary between groups (X 2 = 0.88, pvalue = 0.35). However, while patient age, CCI score, lesion size and BMI were found to not vary between groups, a signi cant variation in steroid dosing between groups was present (p = 2.4e-05), with the WBChi group showing a higher mean daily steroid intake compared to those in the lo group (Fig. 6A-E).
Evaluating the distribution of steroid doses given to patients at time of post-surgical CBC, most patients received low doses of steroids less than 6mg a day (Hi: 76.9%; Lo: 89.5%); however, a larger percent of patient from the WBC-lo fully tapered off steroids (Hi: 24.2%; Lo: 65.1%). Nevertheless, assay of patient distribution shows 55.9% of patients across both groups remain on steroids at follow-up prior to initiating therapy (Fig. 6F). Overall, this nding highlights the additional confounding role steroid intake may have on the survival differences seen between CBC load, but also emphasizes the heterogeneity of steroid dosing in patients following GB resection.

PCA Eigenvector Visualization Can Better Uncover Clinical Relationships in Complex Clinical Datasets
With the adoption of the EHR, patient related data has exponentially grown but remains an under-used resource due to the complexity of mining data and uncovering novel associations. It has been commented that such data provides potential for generating new patient strati cation strategies (30). In turn, our unsupervised PCA eigenvector approach may help rectify these challenges by providing easy to interpret visualizations of clinical data relationships. Illustrating this method utility in our study, the vectors measuring radiation metrics showed opposite directionality to features measuring patient age. The directionality of these vectors thus underscores the known clinical management of GB patients whereby older individuals often receive hypo-fractionated radiation dosage due to toxicity (31). In turn, the application of our eigenvector plot to initially detect relationships of patient survival to CBCs illustrates the utility of applying this method in large clinical datasets as a rst pass visualization approach in identifying novel relationships to explore in a clinical study.  (34). These ndings largely parallel our cox regression models which found patient age, ChemoRT treatment regime, and neutrophil count as predictive to OS. Furthermore, it was noted as well that the median survival time in the WBC-hi group was approximately 12 months while the WBC-lo group was 18 months. Although the coverage of patients in our study ranges the past decade, studies have reported the median survival of GB ranging from 12 to 15 months with ChemoRT treatment (35)(36)(37). In turn, it may be suggestive that patients with lower WBC load, particularly with regards to neutrophil count, experience better survival outcomes than just those with neutrophilia declining more rapidly.
Outside of OS however, our data similarly indicated that WBC load plays a potential role in PFS. While less studies have explored the implication of CBCs to progression outcomes, data has suggested lymphopenia as reducing PFS in primary lesions and use of CBCs including platelet, lymphocyte, and eosinophil counts as predictive in measuring performance outcomes in recurrent GB patients (38-39).
Although these CBC-speci c ndings weren't recapitulated in our studies, our current ndings and past research support the notion that CBC measures are prognostic markers in assessing OS and PFS. In turn, initial pretreatment planning may weigh consideration of these metrics to assist in determining which patients should be more carefully followed after treatment.

Fast Tapering of Steroids following Surgery May In uence Long Term Outcomes of Patients
An important consideration that should be highlighted in our study was the evidence that prior to ChemoRT treatment, large distribution in steroid tapering across our study groups was present. Namely, while our study and others have evidenced that WBC prediction is independent of corticosteroid use statistically, the biologic in uence of corticosteroids to both neutrophil and lymphocyte count has been long recognized (40)(41). Alternative studies, like Dubinski et al., have directly implicated that administration of dexamethasone induces leukocytosis which was associated to poor survival (42). Although steroids have been long used to provide supportive therapy, these ndings may suggest the need to set a more consistent standard of fast tapering patients off steroids (43). Our observed WBC counts and PDL-1 measure may be suggestive of which patients in fact need to have steroid doses modi ed due to in uence on immune cell activity. Studies have evidenced that the increased administration of steroids promotes immune cell dysfunction, namely in T-cell compartments, by promoting increased expression of PDL-1 in the microenvironment that advances dysregulation of immune response (44)(45). Although these ndings may complicate the utilization of CBC measures for survival prediction, these observations more importantly underscore the need to better consider the use of steroids for the symptomatic relief of GB patients.

Conclusion
The use of data derived from the electronic health record will remain a powerful resource in spite of the growing popularity of molecular testing due to the challenges in molecular testing deployment. Moreso, challenges in exploring clinical data from the electronic health record have not been a consequence of non-signi cant data, but rather the sheer complexity of both collecting and assaying data to uncover novel research discoveries. The deliverables of our study include our standardized REDCap data collection form for evaluation of GB, unsupervised analysis framework to initially explore clinical datasets through PCA eigenvector visualization, and automated image analysis pipeline for PDL-1 staining. Furthermore, the measure of CBC load at pretreatment for ChemoRT can be applied to identify patients at risk for unfavorable survival due to high WBC load coupled with elevated PDL-1 staining. However, deeper exploration of these relationships and the tapering of steroids are important considerations in future studies and current approaches to managing GB patients.

Declarations
Availability of data and material The datasets used and/or analyzed during the current study are available from the corresponding authors on reasonable request. Dissemination of the applied REDCap form can be found in supplement or supplied from corresponding authors.     Supplementary Files