Prognostic Model Development with Missing Labels

Background:Across medicine, prognostic models are used to estimate patient risk of certain future health outcomes (e.g., cardiovascular or mortality risk). To develop (or train) prognostic models, historic patient-level training data is needed containing both the predictive factors (i.e., features) and the relevant health outcomes (i.e., labels). Sometimes, when the health outcomes are not recorded in structured data, these are first extracted from textual notes using text mining techniques. Because there exist many studies utilizing text mining to obtain outcome data for prognostic model development, our aim is to study the impact of the text mining quality on downstream prognostic model performance. Methods:We conducted a simulation study charting the relationship between text mining quality and prognostic model performance using an illustrative case study about in-hospital mortality prediction in intensive care unit patients. We repeatedly developed and evaluated a prognostic model for in-hospital mortality, using outcome data extracted by multiple text mining models of varying quality. Results:Interestingly, we found in our case study that a relatively low-quality text mining model (F1 score ≈ 0.50) could already be used to train a prognostic model with quite good discrimination (area under the receiver operating characteristic curve of around 0.80). The calibration of the risks estimated by the prognostic model seemed unreliable across the majority of settings, even when text mining models were of relatively high quality (F1 ≈ 0.80). Discussion:Developing prognostic models on text-extracted outcomes using imperfect text mining models seems promising. However, it is likely that prognostic models developed using this approach may not produce well-calibrated risk estimates, and require recalibration in (possibly a smaller amount of) manually extracted outcome data.

BackgroundIgnoring treatments in prognostic model development or validation can affect the accuracy and transportability of models. We aim to quantify the extent to which the effects of treatment have been addressed in existing prognostic model research and provide recommendations for the handling and reporting of treatment use in future studies.MethodsWe first describe how and when the use of treatments by individuals in a prognostic study can influence the development or validation of a prognostic model. We subsequently conducted a systematic review of the handling and reporting of treatment use in prognostic model studies in cardiovascular medicine. Data on treatment use (e.g. medications, surgeries, lifestyle interventions), the timing of their use, and the handling of such treatment use in the analyses were extracted and summarised.ResultsThree hundred two articles were included in the review. Treatment use was not mentioned in 91 (30%) articles. One hundred forty-six (48%) reported specific information about treatment use in their studies; 78 (26%) provided information about multiple treatments. Three articles (1%) reported changes in medication use (“treatment drop-in”) during follow-up. Seventy-nine articles (26%) excluded treated individuals from their analysis, 80 articles (26%) modelled treatment as an outcome, and of the 155 articles that developed a model, 86 (55%) modelled treatment use, almost exclusively at baseline, as a predictor.ConclusionsThe use of treatments has been partly considered by the majority of CVD prognostic model studies. Detailed accounts including, for example, information on treatment drop-in were rare. Where relevant, the use of treatments should be considered in the analysis of prognostic model studies, particularly when a prognostic model is designed to guide the use of certain treatments and these treatments have been used by the study participants. Future prognostic model studies should clearly report the use of treatments by study participants and consider the potential impact of treatment use on the study findings.

BackgroundLeptin has been implicated in the prognosis of sepsis, yet its mechanistic role remains unclear. This study aimed to develop leptin-associated diagnostic and prognostic models for sepsis and identify potential biomarkers using machine learning approaches.MethodsNon-negative matrix factorization (NMF) was used to identify leptin-related molecular subtypes of sepsis. Weighted gene co-expression network analysis (WGCNA) determined relevant gene modules and hub genes. Differentially expressed genes (DEGs) between sepsis patients and controls were intersected with WGCNA results to refine key genes. Based on these analyses, a prognostic classification model predicting 28-day mortality was developed using the Least Absolute Shrinkage and Selection Operator and Random Forest algorithms, while a time-to-event prognostic model was constructed with Random Survival Forest and Gradient Boosting Machine. Single-cell RNA sequencing was performed to assess expression patterns of core genes across immune cell types. Expression validation was conducted using qPCR and Western blotting.ResultsThree leptin-associated sepsis subtypes with distinct prognoses were identified. The pink and salmon modules from WGCNA were significantly associated with sepsis. Seventy core genes were selected from the DEGs and WGCNA intersection. The prognostic classification model and the time-to-event prognostic model demonstrated strong predictive performance in both the training and external validation cohorts. TFRC and PILRA were consistently highlighted through machine learning, single-cell data, and experimental validation as potential biomarkers.ConclusionWe established leptin-related prognostic models for sepsis using integrated machine learning. TFRC and PILRA may serve as promising biomarkers, offering insights into sepsis heterogeneity and clinical management.

Liver metastasis (LM) is a leading cause of mortality in colorectal cancer (CRC), and currently, no effective therapeutic agents are available. Chanling Gao (CLG) has exhibited inhibitory effects on colorectal cancer liver metastasis (CRLM); however, its exact mechanisms of action remain unclear. The integration of artificial intelligence (AI) with precision traditional Chinese medicine offers a promising approach to enhance therapeutic strategies for CRLM. This study aims to establish a "diagnosis-prognosis" model for CRC and CRLM utilizing AI, and to explore the potential mechanisms of CLG treatment for CRC and CRLM through biological information analysis and cellular experiments. A "component-target" network was constructed for elucidating the mechanisms underpinning the therapeutic potential of CLG in CRLM through network pharmacology. Prognostic models for CRC were developed by combining Non-Negative Matrix Factorization (NMF) clustering with ten machine learning (ML) methods, using core targets identified from the network, and validated across multiple TCGA and GEO cohorts. Clinical pathological factors, survival data, biological functional enrichment, and immune landscape analyses were performed to construct a nomogram and compare the results with those of previously published studies. A diagnostic model for CRLM was developed by employing ten ML techniques in cross-combination using genes from the prognostic model, followed by an analysis of immune microenvironmental differences between CRC and CRLM at the single-cell and spatial transcriptome levels. Key targets involved in CRC onset and CRLM progression were identified, and a transcription factor (TF) regulatory network was established by screening the upstream TFs of these targets. Fourteen phenotypic functions were scored to determine their associations with key targets. Molecular docking and dynamics simulations were performed to assess the binding affinity of CLG components with key targets (TP53, CDK1, and CCNB1). In vitro cell experiments verified the inhibitory effects of CLG and its components on colorectal cancer and their regulatory roles on critical targets. The "active ingredient-target" network for CLG identified 248 intersecting targets. NMF clustering revealed two prognostic subtypes, i.e., C1 and C2, with C1 demonstrating superior prognostic outcomes over C2. Survival outcomes and immune differences between the subtypes were analyzed, resulting in the identification of 47 core targets by intersecting differentially expressed genes with previously identified targets. A CLG risk score-based prognostic ML model was constructed using ten ML cross-combination approaches and validated for survival prognosis, clinical diagnosis, and therapeutic utility across TCGA and GEO cohorts. Immune landscape analysis revealed that low-risk groups were characterized by a "hot tumor" phenotype with a favorable response to immunotherapy, whereas high-risk groups exhibited a "cold tumor" phenotype, with potential immunotherapy benefits. Independent prognostic analysis confirmed the CLG-derived risk score independently predicted prognosis. Validation against ten published models demonstrated elevated accuracy and efficacy for the proposed model.A CRLM diagnostic model was constructed using 11 genes from the prognostic model. Receiver operating characteristic (ROC) curve, calibration plot, and decision curve (DCA) analyses demonstrated its accuracy and clinical utility, indicating high predictive efficiency. Immune microenvironmental differences identified CDK1 and CCNB1 as potential biomarkers associated with CRLM onset and progression. CDK1 and CCNB1 expression levels had positive correlations with M2 macrophages, known to promote liver metastasis. The TF regulatory network revealed a regulatory relationship involving TP53, CDK1, and CCNB1, while gene set variation analysis (GSVA) demonstrated the associations of CDK1/CCNB1 with cell cycle regulation and apoptosis. Molecular docking and dynamics simulations revealed strong binding affinities of CLG components with key targets. In vitro experiments confirmed that CLG and its components effectively inhibit colorectal cancer and regulate critical gene expression. This study established an "active ingredient-target" network for CLG using network pharmacology and developed a dynamic "diagnosis-prognosis "model integrating ML based on drug targets. The clinical value of the model in CRC and CRLM patients was validated, and drug-target binding was elucidated using DL. It was discovered that CLG may inhibit CRC liver metastasis by targeting the TP53/CCNB1/CDK1 signaling pathway. These findings highlight the clinical utility of CLG and dynamic models in the prevention, diagnosis, and therapeutic management of CRC and CRLM, providing a robust foundation in bioinformatics, pharmacology, and potential targets for CRLM treatment in TCM, with broad clinical implications.

PurposeTo identify risk factors and develop a prediction model for the development of profound and recurrent shock amongst children presenting with dengue shock syndrome (DSS)MethodsWe analyzed data from a prospective cohort of children with DSS recruited at the Paediatric Intensive Care Unit of the Hospital for Tropical Disease in Ho Chi Minh City, Vietnam. The primary endpoint was “profound DSS”, defined as ≥2 recurrent shock episodes (for subjects presenting in compensated shock), or ≥1 recurrent shock episodes (for subjects presenting initially with decompensated/hypotensive shock), and/or requirement for inotropic support. Recurrent shock was evaluated as a secondary endpoint. Risk factors were pre-defined clinical and laboratory variables collected at the time of presentation with shock. Prognostic model development was based on logistic regression and compared to several alternative approaches.ResultsThe analysis population included 1207 children of whom 222 (18%) progressed to “profound DSS” and 433 (36%) had recurrent shock. Independent risk factors for both endpoints included younger age, earlier presentation, higher pulse rate, higher temperature, higher haematocrit and, for females, worse hemodynamic status at presentation. The final prognostic model for “profound DSS” showed acceptable discrimination (AUC=0.69 for internal validation) and calibration and is presented as a simple score-chart.ConclusionsSeveral risk factors for development of profound or recurrent shock among children presenting with DSS were identified. The score-chart derived from the prognostic models should improve triage and management of children presenting with DSS in dengue-endemic areas.

This study presents a novel approach for predicting hierarchical short time series. In this article, our objective was to formulate long-term forecasts for household natural gas consumption by considering the hierarchical structure of territorial units within a country’s administrative divisions. For this purpose, we utilized natural gas consumption data from Poland. The length of the time series was an important determinant of the data set. We contrast global techniques, which employ a uniform method across all time series, with local methods that fit a distinct method for each time series. Furthermore, we compare the conventional statistical approach with a machine learning (ML) approach. Based on our analyses, we devised forecasting methods for short time series that exhibit exceptional performance. We have demonstrated that global models provide better forecasts than local models. Among ML models, neural networks yielded the best results, with the MLP network achieving comparable performance to the LSTM network while requiring significantly less computational time.

BackgroundRather low vaccination rates for Human papillomavirus (HPV) and pre-existing cervical cancer patients with limited therapeutic strategies ask for more precise prognostic model development. On the other side, the clinical significance of circadian clock signatures in cervical cancer lacks investigation.MethodsSubtypes classification based upon eight circadian clock core genes were implemented in TCGA-CESC through k-means clustering methods. Afterwards, KEGG, GO and GSEA analysis were conducted upon differentially expressed genes (DEGs) between high and low-risk groups, and tumor microenvironment (TME) investigation by CIBERSORT and ESTIMATE. Furthermore, a prognostic model was developed by cox and lasso regression methods, and verified in GSE44001 by time-dependent receiver-operating characteristic curve (ROC) analysis. Lastly, FISH and IHC were used for validation of CCL20 expression in patients’ specimens and U14 subcutaneous tumor models were built for TME composition.ResultsWe successfully classified cervical patients into high-risk and low-risk groups based upon circadian-oscillation-signatures. Afterwards, we built a prognostic risk model composed of GJB2, CCL20 and KRT24 with excellent predictive value on patients’ overall survival (OS). We then proposed metabolism unbalance, especially for glycolysis, and immune related pathways to be major enriched signatures between the high-risk and low-risk groups. Then, we proposed an ‘immune-desert’-like suppressive myeloid cells infiltration pattern in high-risk group TME and verified its resistance to immunotherapies. Finally, CCL20 was proved positively correlated with real-world patients’ stages and induced significant less CD8+ T cells and more M2 macrophages infiltration in mouse model.ConclusionsWe unraveled a prognostic risk model based upon circadian oscillation and verified its solidity. Specifically, we unveiled distinct TME immune signatures in high-risk groups.

Recent studies have unveiled disulfidptosis as a phenomenon intimately associated with cellular damage, heralding new avenues for exploring tumor cell dynamics. We aimed to explore the impact of disulfide cell death on the tumor immune microenvironment and immunotherapy in lung adenocarcinoma (LUAD). We initially utilized pan-cancer transcriptomics to explore the expression, prognosis, and mutation status of genes related to disulfidptosis. Using the LUAD multi- -omics cohorts in the TCGA database, we explore the molecular characteristics of subtypes related to disulfidptosis. Employing various machine learning algorithms, we construct a robust prognostic model to predict immune therapy responses and explore the model's impact on the tumor microenvironment through single-cell transcriptome data. Finally, the biological functions of genes related to the prognostic model are verified through laboratory experiments. Genes related to disulfidptosis exhibit high expression and significant prognostic value in various cancers, including LUAD. Two disulfidptosis subtypes with distinct prognoses and molecular characteristics have been identified, leading to the development of a robust DSRS prognostic model, where a lower risk score correlates with a higher response rate to immunotherapy and a better patient prognosis. NAPSA, a critical gene in the risk model, was found to inhibit the proliferation and migration of LUAD cells. Our research introduces an innovative prognostic risk model predicated upon disulfidptosis genes for patients afflicted with Lung Adenocarcinoma (LUAD). This model proficiently forecasts the survival rates and therapeutic outcomes for LUAD patients, thereby delineating the high-risk population with distinctive immune cell infiltration and a state of immunosuppression. Furthermore, NAPSA can inhibit the proliferation and invasion capabilities of LUAD cells, thereby identifying new molecules for clinical targeted therapy.

Tumor-associated macrophages (TAMs) are crucial in hepatocellular carcinoma (HCC) development and invasion. This study explores monocyte/ macrophage-associated gene expression profiles in HCC, constructs a prognostic model based on these genes, and examines its relationship with drug resistance and immune therapy responses. Single-cell RNA sequencing(scRNA-seq) data from 10 HCC tissue biopsy samples, totaling 24,597 cells, were obtained from the GEO database to identify monocyte/macrophage-associated genes. A prognostic model was constructed and validated using external datasets and Western blot. Relationships between the model, clinical correlates, drug sensitivity, and immune therapy responses were investigated. From scRNA-seq data, 2,799 monocyte/macrophage marker genes were identified. Using the TCGA dataset, a prognostic model based on the single-gene UQCRH was constructed, stratifying patients into high-risk and low-risk groups based on overall survival rates. High-risk group patients showed reduced survival rates and higher UQCRH expression in tumor tissues. Western blot analysis further confirmed the elevated expression of UQCRH in HCC cell lines. Spatial transcriptomics analysis revealed that high UQCRH expression co-localized with malignant cells in the tumor tissue. Drug sensitivity analysis revealed that the high-risk group had lower sensitivity to sorafenib and axitinib. Immune therapy response analysis indicated poorer outcomes in the high-risk group, with more pronounced APC inhibition and a weaker IFN-II response. Clinical indicator analysis showed a positive correlation between high UQCRH expression and tumor invasion. Enrichment analysis of UQCRH and associated molecules indicated involvement in oxidative phosphorylation and mitochondrial electron transport. This study introduces a prognostic model for HCC patients based on monocyte/macrophage marker genes. The single-gene model predicts HCC patient survival and treatment outcomes, identifying high-risk individuals with varying drug sensitivities and immune suppression states.

Clear cell renal cell carcinoma's (ccRCC) occurrence and development are strongly linked to the metabolic reprogramming of tumors, and thus far, neither its prognosis nor treatment has achieved satisfying clinical outcomes. The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases, respectively, provided us with information on the RNA expression of ccRCC patients and their clinical data. Cuproptosis-related genes (CRGS) were discovered in recent massive research. With the help of log-rank testing and univariate Cox analysis, the prognostic significance of CRGS was examined. Different cuproptosis subtypes were identified using consensus clustering analysis, and GSVA was used to further investigate the likely signaling pathways between various subtypes. Univariate Cox, least absolute shrinkage and selection operator (Lasso), random forest (RF), and multivariate stepwise Cox regression analysis were used to build prognostic models. After that, the models were verified by means of the C index, Kaplan-Meier (K-M) survival curves, and time-dependent receiver operating characteristic (ROC) curves. The association between prognostic models and the tumor immune microenvironment as well as the relationship between prognostic models and immunotherapy were next examined using ssGSEA and TIDE analysis. Four online prediction websites-Mircode, MiRDB, MiRTarBase, and TargetScan-were used to build a lncRNA-miRNA-mRNA ceRNA network. By consensus clustering, two subgroups of cuproptosis were identified that represented distinct prognostic and immunological microenvironments. A prognostic risk model with 13 CR-lncRNAs was developed. The immune microenvironment and responsiveness to immunotherapy are substantially connected with the model, which may reliably predict the prognosis of patients with ccRCC.

<h3>Abstract</h3> <h3>Background</h3> COVID19 is worldwide pandemic that is mild in the majority of patients but can result in a pneumonia like illness with progression to acute respiratory distress syndrome and death. Predicting the disease severity at time of diagnosis can be helpful in prioritizing hospital admission and resources. <h3>Methods</h3> We prospectively recruited 1096 consecutive patients with COVID19 from the Jaber Hospital, a COVID19 facility in Kuwait, between 24 February and 20 April 2020. The primary endpoint of interest was disease severity defined algorithmically. Predefined risk variables were collected at the time of PCR based diagnosis of the infection. Prognostic model development used 5-fold cross-validated regularized logit regression. The cohort was divided into a training and validation cohort and all model development proceeded on the training cohort. <h3>Results</h3> There were 643 patients with clinical course data of whom 94 developed severe COVID19. In the final model, age, CRP, procalcitonin, lymphocyte and monocyte percentages and serum albumin were independent predictors of a more severe illness course. The final prognostic model demonstrated good discrimination, calibration and internal validity. <h3>Conclusion</h3> We developed and validated a simple score calculated at time of diagnosis that can predict patients with severe COVID19 disease.

A fundamental part of medical research is the development and validation of diagnostic and prognostic prediction models [1,2]. These prediction models aim to predict the absolute probability that a certain disease or condition is currently present (diagnostic models) or that an outcome will occur within a specific follow-up period (prognostic models) for an individual subject. Prediction models typically rely on multiple predictors, which can include demographic characteristics, medical history and physical examination items, or more complex measurements from, for example, medical imaging, electrophysiology, pathology, and biomarkers. Also for diagnostic models, estimates of probabilities are rarely based on a single test, and doctors naturally integrate several patient characteristics and symptoms [3]. A broad range of prediction modeling techniques exist, like regression approaches, neural network models, decision tree models, genetic programming models, and support vector machine learning models, although prediction models developed by a multivariable regression approach are by far prevailing. It is widely recommended that a developed prediction model should not be used in practice before being externally validated—at least once—in other individuals than those used for model development [4–7]. Unfortunately, most prediction models are poorly or not at all validated, rendering interpretation of their generalizability difficult. In addition, many systematic reviews showed that for the same outcome or same target population, numerous competing models exist [8–10]. Generally speaking, researchers often ignore existing prediction models and develop yet another prediction model from their own data [2]. This practice sustains a cycle of underpowered prediction model development studies and poor knowledge about the generalizability and applicability of developed prediction models. Evidence synthesis and meta-analysis of individual participant data (IPD) from multiple studies seems to be a unique opportunity to address these problems, as it allows researchers to develop and directly validate models on large datasets and across a wide range of populations and settings, to directly test a model’s generalizability (Fig 1) [11–13]. Fig 1 Trends in publications of IPD-MA studies focusing on the development and/or validation of diagnostic or prognostic prediction models. There is currently little guidance on how to conduct an IPD meta-analysis (IPD-MA) for developing and/or validating diagnostic or prognostic prediction models [15]. To date, most IPD-MA articles focus on estimating relative quantities, like a risk ratio, hazard ratio, or odds ratio for a specific treatment or a specific etiologic factor. In contrast, prediction modeling research is focused on developing and validating multivariable models aimed at calculating an absolute risk estimate of the combined variables, rather than estimating the relative effect of a specific treatment or etiologic factor. Furthermore, prediction modeling studies focus entirely on the role and joint contribution of multiple covariates, whereas intervention studies in principle rely on randomization to reduce the role of covariates (Table 1). Hence, IPD-MAs of randomized intervention and etiological studies, which are beyond the scope of this paper and are instead addressed in the accompanying paper [16], differ from IPD-MAs of multivariable prediction models, which are the focus of this paper. Table 1 The main differences between IPD-MA of treatment intervention studies and of multivariable prediction modeling studies. We provide an overview of the advantages and limitations of IPD-MAs aiming to develop a novel prediction model or to validate one or more existing models across multiple datasets. This overview is based on published guidelines and existing recommendations for the conduct of prediction modeling studies and of IPD-MA research. We illustrate this overview with examples of recently published IPD-MAs of prediction models across various medical domains. Our aim is to help researchers, readers, reviewers, and editors to identify and understand the key issues involved with such IPD-MA projects.

Both immune profiling and tumor budding significantly correlate with colorectal cancer patient outcome but are traditionally reported independently. This study evaluated the association and interaction between lymphocytic infiltration and tumor budding, coregistered on a single slide, in order to determine a more precise prognostic algorithm for patients with stage II colorectal cancer. Multiplexed immunofluorescence and automated image analysis were used for the quantification of CD3+CD8+ T cells, and tumor buds (TBs), across whole slide images of three independent cohorts (training cohort: n = 114, validation cohort 1: n = 56, validation cohort 2: n = 62). Machine learning algorithms were used for feature selection and prognostic risk model development. High numbers of TBs [HR = 5.899; 95% confidence interval (CI) 1.875-18.55], low CD3+ T-cell density (HR = 9.964; 95% CI, 3.156-31.46), and low mean number of CD3+CD8+ T cells within 50 μm of TBs (HR = 8.907; 95% CI, 2.834-28.0) were associated with reduced disease-specific survival. A prognostic signature, derived from integrating TBs, lymphocyte infiltration, and their spatial relationship, reported a more significant cohort stratification (HR = 18.75; 95% CI, 6.46-54.43), than TBs, Immunoscore, or pT stage. This was confirmed in two independent validation cohorts (HR = 12.27; 95% CI, 3.524-42.73; HR = 15.61; 95% CI, 4.692-51.91). The investigation of the spatial relationship between lymphocytes and TBs within the tumor microenvironment improves accuracy of prognosis of patients with stage II colorectal cancer through an automated image analysis and machine learning workflow.

ObjectiveBreast cancer (BC) remains the most prevalent malignancy among women. Clinical evidence indicates that genetic variations related to circadian rhythms, as well as the timing of therapeutic interventions, influence the response to radiation therapy and the toxicity of pharmacological treatments in women with BC. This study aimed to identify key circadian rhythm-related genes (CRGs) using bioinformatics and machine learning, and construct a prognostic model to predict clinical outcomes.MethodsTranscriptome data for BC were retrieved from The Cancer Genome Atlas database. Univariate Cox regression and least absolute shrinkage and selection operator regression analyses were used to develop a prognostic model based on CRGs. The predictive performance of the risk score model was evaluated. Univariate and multivariate Cox regression analyses were applied to construct the prognostic model and stratify patients into high-risk and low-risk groups. Additionally, differences in immune microenvironment, immunotherapy efficacy, and tumor mutation burden were assessed between risk groups.ResultsA prognostic risk score model comprising 17 CRGs was developed. The areas under the receiver operating characteristic curve for overall survival at 1, 3, 5, and 7 years exceeded 0.6, indicating acceptable predictive performance. Calibration plots and decision curve analyses demonstrated the use of the model in prognostic prediction. Significant differences in immune microenvironment, immunotherapy efficacy, and tumor mutation burden were identified between the low-risk and high-risk groups.ConclusionThe circadian rhythm-based gene model, effectively predicted the prognosis of individuals with BC, highlighting its potential to inform personalized therapeutic strategies and improve patient outcomes.

Traumatic brain injury (TBI) is a leading cause of death and disability worldwide, with a disproportionate burden of this injury on low- and middle-income countries (LMICs). Limited access to diagnostic technologies and highly skilled providers combined with high patient volumes contributes to poor outcomes in LMICs. Prognostic modeling as a clinical decision support tool, in theory, could optimize the use of existing resources and support timely treatment decisions in LMICs. The objective of this study was to develop a machine learning-based prognostic model using data from Kilimanjaro Christian Medical Centre in Moshi, Tanzania. This study is a secondary analysis of a TBI data registry including 3138 patients. The authors tested nine different machine learning techniques to identify the prognostic model with the greatest area under the receiver operating characteristic curve (AUC). Input data included demographics, vital signs, injury type, and treatment received. The outcome variable was the discharge score on the Glasgow Outcome Scale-Extended. The AUC for the prognostic models varied from 66.2% (k-nearest neighbors) to 86.5% (Bayesian generalized linear model). An increasing Glasgow Coma Scale score, increasing pulse oximetry values, and undergoing TBI surgery were predictive of a good recovery, while injuries suffered from a motor vehicle crash and increasing age were predictive of a poor recovery. The authors developed a TBI prognostic model with a substantial level of accuracy in a low-resource setting. Further research is needed to externally validate the model and test the algorithm as a clinical decision support tool.