Predictions For Individual Patients Research Articles

Creatinine clearance/eGFR ratio: a simple index for muscle mass related to mortality in ICU patients

IntroductionIn patients admitted to the intensive care unit (ICU), muscle mass is inversely associated with mortality. Although muscle mass can be estimated with 24-h urinary creatinine excretion (UCE), its use for risk prediction in individual patients is limited because age-, sex-, weight- and length-specific reference values for UCE are lacking. The ratio between measured creatinine clearance (mCC) and estimated glomerular filtration rate (eGFR) might circumvent this constraint. The main goal was to assess the association of the mCC/eGFR ratio in ICU patients with all-cause hospital and long-term mortality.MethodsThe mCC/eGFR ratio was determined in patients admitted to our ICU between 2005 and 2021 with KDIGO acute kidney injury (AKI) stage 0–2 and an ICU stay ≥ 24 h. mCC was calculated from UCE and plasma creatinine and indexed to 1.73 m2. mCC/eGFR was analyzed by categorizing patients in mCC/eGFR quartiles and as continuous variable.ResultsSeven thousand five hundred nine patients (mean age 61 ± 15 years; 38% female) were included. In-hospital mortality was 27% in the lowest mCC/eGFR quartile compared to 11% in the highest quartile (P < 0.001). Five-year post-hospital discharge actuarial mortality was 37% in the lowest mCC/eGFR quartile compared to 19% in the highest quartile (P < 0.001). mCC/eGFR ratio as continuous variable was independently associated with in-hospital mortality in multivariable logistic regression (odds ratio: 0.578 (95% CI: 0.465—0.719); P < 0.001). mCC/eGFR ratio as continuous variable was also significantly associated with 5-year post-hospital discharge mortality in Cox regression (hazard ratio: 0.27 (95% CI: 0.22—0.32); P < 0.001).ConclusionsThe mCC/eGFR ratio is associated with both in-hospital and long-term mortality and may be an easily available index of muscle mass in ICU patients.

Electroconvulsive therapy response and remission in moderate to severe depressive illness: a decade of national Scottish data.

Despite strong evidence of efficacy of electroconvulsive therapy (ECT) in the treatment of depression, no sensitive and specific predictors of ECT response have been identified. Previous meta-analyses have suggested some pre-treatment associations with response at a population level. Using 10 years (2009-2018) of routinely collected Scottish data of people with moderate to severe depression (n = 2074) receiving ECT we tested two hypotheses: (a) that there were significant group-level associations between post-ECT clinical outcomes and pre-ECT clinical variables and (b) that it was possible to develop a method for predicting illness remission for individual patients using machine learning. Data were analysed on a group level using descriptive statistics and association analyses as well as using individual patient prediction with machine learning methodologies, including cross-validation. ECT is highly effective for moderate to severe depression, with a response rate of 73% and remission rate of 51%. ECT response is associated with older age, psychotic symptoms, necessity for urgent intervention, severe distress, psychomotor retardation, previous good response, lack of medication resistance, and consent status. Remission has the same associations except for necessity for urgent intervention and, in addition, history of recurrent depression and low suicide risk. It is possible to predict remission with ECT with an accuracy of 61%. Pre-ECT clinical variables are associated with both response and remission and can help predict individual response to ECT. This predictive tool could inform shared decision-making, prevent the unnecessary use of ECT when it is unlikely to be beneficial and ensure prompt use of ECT when it is likely to be effective.

Novel machine-learning prediction tools for overall survival of patients with chondrosarcoma: Based on recursive partitioning analysis.

Chondrosarcoma (CHS), a bone malignancy, poses a significant challenge due to its heterogeneous nature and resistance to conventional treatments. There is a clear need for advanced prognostic instruments that can integrate multiple prognostic factors to deliver personalized survival predictions for individual patients. This study aimed to develop a novel prediction tool based on recursive partitioning analysis (RPA) to improve the estimation of overall survival for patients with CHS. Data from the Surveillance, Epidemiology, and End Results (SEER) database were analyzed, including demographic, clinical, and treatment details of patients diagnosed between 2000 and 2018. Using C5.0 algorithm, decision trees were created to predict survival probabilities at 12, 24, 60, and 120 months. The performance of the models was assessed through confusion scatter plot, accuracy rate, receiver operator characteristic (ROC) curve, and area under ROC curve (AUC). The study identified tumor histology, surgery, age, visceral (brain/liver/lung) metastasis, chemotherapy, tumor grade, and sex as critical predictors. Decision trees revealed distinct patterns for survival prediction at each time point. The models showed high accuracy (82.40%-89.09% in training group, and 82.16%-88.74% in test group) and discriminatory power (AUC: 0.806-0.894 in training group, and 0.808-0.882 in test group) in both training and testing datasets. An interactive web-based shiny APP (URL: https://yangxg1209.shinyapps.io/chondrosarcoma_survival_prediction/) was developed, simplifying the survival prediction process for clinicians. This study successfully employed RPA to develop a user-friendly tool for personalized survival predictions in CHS. The decision tree models demonstrated robust predictive capabilities, with the interactive application facilitating clinical decision-making. Future prospective studies are recommended to validate these findings and further refine the predictive model.

Depression-Anxiety Coupling Strength as a predictor of relapse in major depressive disorder: A CAN-BIND wellness monitoring study report

BackgroundA critical challenge in the study and management of major depressive disorder (MDD) is predicting relapse. We examined the temporal correlation/coupling between depression and anxiety (called Depression-Anxiety Coupling Strength, DACS) as a predictor of relapse in patients with MDD. MethodsWe followed 97 patients with remitted MDD for an average of 394 days. Patients completed weekly self-ratings of depression and anxiety symptoms using the Quick Inventory of Depressive Symptoms (QIDS-SR) and the Generalized Anxiety Disorder 7-item scale (GAD-7). Using these longitudinal ratings we computed DACS as random slopes in a linear mixed effects model reflecting individual-specific degree of correlation between depression and anxiety across time points. We then tested DACS as an independent variable in a Cox proportional hazards model to predict relapse. ResultsA total of 28 patients (29 %) relapsed during the follow-up period. DACS significantly predicted confirmed relapse (hazard ratio [HR] 1.5, 95 % CI [1.01, 2.22], p = 0.043; Concordance 0.79 [SE 0.04]). This effect was independent of baseline depressive or anxiety symptoms or their average levels over the follow-up period, and was identifiable more than one month before relapse onset. LimitationsSmall sample size, in a single study. Narrow phenotype and comorbidity profiles. ConclusionsDACS may offer opportunities for developing novel strategies for personalized monitoring, early detection, and intervention. Future studies should replicate our findings in larger, diverse patient populations, develop individual patient prediction models, and explore the underlying mechanisms that govern the relationship of DACS and relapse.

Model-Informed Precision Dosing Using Machine Learning for Levothyroxine in General Practice: Development, Validation and Clinical Simulation Trial.

Levothyroxine is one of the most prescribed drugs in the western world. Dosing is challenging due to high-interindividual differences in effective dosage and the narrow therapeutic window. Model-informed precision dosing (MIPD) using machine learning could assist general practitioners (GPs), but no such models exist for primary care. Furthermore, introduction of decision-support algorithms in healthcare is limited due to the substantial gap between developers and clinicians' perspectives. We report the development, validation, and a clinical simulation trial of the first MIPD application for primary care. Stable maintenance dosage of levothyroxine was the model target. The multiclass model generates predictions for individual patients, for different dosing classes. Random forest was trained and tested on a national primary care database (n = 19,004) with a final weighted AUC across dosing options of 0.71, even in subclinical hypothyroidism. TSH, fT4, weight, and age were most predictive. To assess the safety, feasibility, and clinical impact of MIPD for levothyroxine, we performed clinical simulation studies in GPs and compared MIPD to traditional prescription. Fifty-one GPs selected starting dosages for 20 primary hypothyroidism cases without and then with MIPD 2 weeks later. Overdosage and underdosage were defined as higher and lower than 12.5 μg relative to stable maintenance dosage. MIPD decreased overdosage in number (30.5 to 23.9%, P < 0.01) and magnitude (median 50 to 37.5 μg, P < 0.01) and increased optimal starting dosages (18.3 to 30.2%, P < 0.01). GPs considered lab results more often with MIPD and most would use the model frequently. This study demonstrates the clinical relevance, safety, and effectiveness of MIPD for levothyroxine in primary care.

Development and validation of machine learning models to predict postoperative infarction in moyamoya disease.

Cerebral infarction is a common complication in patients undergoing revascularization surgery for moyamoya disease (MMD). Although previous statistical evaluations have identified several risk factors for postoperative brain ischemia, the ability to predict its occurrence based on these limited predictors remains inadequately explored. This study aimed to assess the feasibility of machine learning algorithms for predicting cerebral infarction after revascularization surgery in patients with MMD. This retrospective study was conducted across two centers and harnessed data from 512 patients with MMD who had undergone revascularization surgery. The patient cohort was partitioned into internal and external datasets. Using perioperative clinical data from the internal cohort, three distinct machine learning algorithms-namely the support vector machine, random forest, and light gradient-boosting machine models-were trained and cross-validated to predict the occurrence of postoperative cerebral infarction. Predictive performance validity was subsequently assessed using an external dataset. Shapley additive explanations (SHAP) analysis was conducted to augment the prediction model's transparency and to quantify the impact of each input variable on shaping both the aggregate and individual patient predictions. In the cohort of 512 patients, 33 (6.4%) experienced postrevascularization cerebral infarction. The cross-validation outcomes revealed that, among the three models, the support vector machine model achieved the largest area under the receiver operating characteristic curve (ROC-AUC) at mean ± SD 0.785 ± 0.052. Notably, during external validation, the light gradient-boosting machine model exhibited the highest accuracy at 0.903 and the largest ROC-AUC at 0.710. The top-performing prediction model utilized five input variables: postoperative serum gamma-glutamyl transpeptidase value, positive posterior cerebral artery (PCA) involvement on preoperative MRA, infarction as the rationale for surgery, presence of an infarction scar on preoperative MRI, and preoperative modified Rankin Scale score. Furthermore, the SHAP analysis identified presence of PCA involvement, infarction as the rationale for surgery, and presence of an infarction scar on preoperative MRI as positive influences on postoperative cerebral infarction. This study indicates the usefulness of employing machine learning techniques with routine perioperative data to predict the occurrence of cerebral infarction after revascularization procedures in patients with MMD.

Abstract 3886: Analysis of clonal heterogeneity within paired primary and metastatic tumor samples of patients with solid tumors and implications for neoantigen-based personalized cancer vaccines

Abstract Background: Neoantigen-based personalized cancer vaccines carry significant promise in treating solid malignancies. However, there are uncertainties regarding the choice between the primary or the metastatic tumor for neoantigen prediction in individual patients. Here, we conducted a thorough examination of somatic variations in 676 patients who had paired primary and metastatic solid tumors. Methods: Patients were enrolled in the Total Cancer Care protocol (NCT03977402) to which patients provided an IRB-approved written informed consent within the Oncology Research Information Exchange Network (ORIEN). Whole-exome sequencing of 756 primary and metastatic tumor pairs was performed (N = 676 patients). These included Genitourinary (n=83), Gynecological (n=97), Gastrointestinal (n=213), Thoracic (n=33), Cutaneous (n=24), Breast (n=108), Endometrial (n=49), Sarcoma (n=35), Head-and-Neck (n=106) and others (n=8). The data was analyzed through the ORIEN AVATAR Molecular Analysis Pipeline for somatic mutation variant detection and variant annotation. In this analysis, we focused on somatic events that result in an in-frame alteration (such as missense, in-frame deletion and in-frame insertion) and out-of-frame protein-altering mutations (such as frameshifts, de novo start, out-of-frame, and nonstop gain). Clonal population structure was determined based on pyclone-vi. Results: For in-frame events, bladder cancer, melanoma, and gynecological cancers shared close to 75% of the mutations between paired primary and metastatic cases. In contrast, sarcoma and thyroid cancer had a low overlap (~ 25%) of variants. For out-of-frame events, these events tend to have a lower proportion of shared somatic variants between primary and metastasis than in-frame variants. Oncogenic drivers (e.g., BRAF V600E, KRAS G12A, and TP53 loss-of-function) were highly likely to be present in both paired primary and metastatic tumors. Next, we performed additional analysis on evolutionary selection of protein-coding variants via dN/dS calculation. We found no significant global shift in dN/dS ratio between paired primary and metastatic tumors across malignancies. However, we found increased selection of protein-coding variants in brain and liver metastatic sites, which correlated with increased homologous recombination deficiency within these sites. Conclusions: Our analysis demonstrates genetic variations that exist when comparing paired primary and metastatic tumors that appear to vary by histology. Variants are potentially undergoing negative selection supported by the preferential loss of out-of-frame events in metastatic tumors and positive selection in specific metastatic sites. Overall, understanding the clonal structure will be key to neoantigen prediction for effective neoantigen-based vaccines. Citation Format: Alyssa Obermayer, Timothy Shaw, Darwin Chang, Joshua Davis, Jamie K. Teer, Xiaoqing Yu, Xuefeng Wang, Dale Hedges, Aik Choon Tan, Robert Rounbehler, Abdul Rafeh Naqash, Margaret Gatti-Mays, Aakrosh Ratan, Martin McCarter, Howard Colman, Igor Puzanov, Susanne Arnold, Michelle Churchman, Patrick Hwu, William Dalton, George Weiner, Jose Conejo-Garcia, Paulo C. Rodriguez, Bodour Salhia, Ahmad A. Tarhini. Analysis of clonal heterogeneity within paired primary and metastatic tumor samples of patients with solid tumors and implications for neoantigen-based personalized cancer vaccines [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 3886.

An updated PREDICT breast cancer prognostic model including the benefits and harms of radiotherapy

PREDICT Breast (www.breast .predict.nhs.uk) is a prognostication tool for early invasive breast cancer. The current version was based on cases diagnosed in 1999–2003 and did not incorporate the benefits of radiotherapy or the harms associated with therapy. Since then, there has been a substantial improvement in the outcomes for breast cancer cases. The aim of this study was to update PREDICT Breast to ensure that the underlying model is appropriate for contemporary patients. Data from the England National Cancer Registration and Advisory Service for invasive breast cancer cases diagnosed 2000–17 were used for model development and validation. Model development was based on 35,474 cases diagnosed and registered by the Eastern Cancer Registry. A Cox model was used to estimate the prognostic effects of the year of diagnosis, age at diagnosis, tumour size, tumour grade and number of positive nodes. Separate models were developed for ER-positive and ER-negative disease. Data on 32,408 cases from the West Midlands Cancer Registry and 100,551 cases from other cancer registries were used for validation. The new model was well-calibrated; predicted breast cancer deaths at 5-, 10- and 15-year were within 10 per cent of the observed validation data. Discrimination was also good: The AUC for 15-year breast cancer survival was 0.809 in the West Midlands data set and 0.846 in the data set for the other registries. The new PREDICT Breast model outperformed the current model and will be implemented in the online tool which should lead to more accurate absolute treatment benefit predictions for individual patients.

Cognitive functioning in untreated glioma patients: The limited predictive value of clinical variables.

Previous research identified many clinical variables that are significantly related to cognitive functioning before surgery. It is not clear whether such variables enable accurate prediction for individual patients' cognitive functioning because statistical significance does not guarantee predictive value. Previous studies did not test how well cognitive functioning can be predicted for (yet) untested patients. Furthermore, previous research is limited in that only linear or rank-based methods with small numbers of variables were used. We used various machine learning models to predict preoperative cognitive functioning for 340 patients with glioma across 18 outcome measures. Predictions were made using a comprehensive set of clinical variables as identified from the literature. Model performances and optimized hyperparameters were interpreted. Moreover, Shapley additive explanations were calculated to determine variable importance and explore interaction effects. Best-performing models generally demonstrated above-random performance. Performance, however, was unreliable for 14 out of 18 outcome measures with predictions worse than baseline models for a substantial number of train-test splits. Best-performing models were relatively simple and used most variables for prediction while not relying strongly on any variable. Preoperative cognitive functioning could not be reliably predicted across cognitive tests using the comprehensive set of clinical variables included in the current study. Our results show that a holistic view of an individual patient likely is necessary to explain differences in cognitive functioning. Moreover, they emphasize the need to collect larger cross-center and multimodal data sets.

OP134 Mapping The Health Of The Nation Outcomes Scale To The EQ-5D Using The Pharmacotherapy Monitoring And Outcome Survey

IntroductionIn clinical practice, health outcomes relevant for economic evaluations, such as health-related quality of life (i.e., utilities), are not always available. However, algorithms can be developed to map available outcome measures to utilities. In the Netherlands, the Health of the Nations Outcomes Scores (HoNOS) is a physician-reported, disease-specific outcome measure frequently used for patients with severe mental illness for which no mapping algorithms for utilities currently exist. The aim of this study was to develop an algorithm to map responses on the HoNOS to EQ-5D utilities for use in economic evaluations.MethodsA dataset was obtained from the Pharmacotherapy Monitoring and Outcome Survey cohort study containing data from patients with psychotic disorders. The dataset contains EQ-5D-3L domain and HoNOS scores, the age and sex of patients, and additional demographics. Correlation between the EQ-5D-3L and HoNOS was evaluated. To derive mapping functions, least absolute shrinkage and selection operator (LASSO) regression and random forest algorithms were applied with various predictor variables using a machine learning approach, whereby data were split into separate training and test sets. Cross-validation was then used to compare the performance of different models using R-squared and the root mean square error (RMSE).ResultsA total of 2,111 patients were included in the study. Spearman’s correlation coefficients indicated a weak to moderate negative correlation of -0.31. Based on model performance metrics, LASSO models outperformed random forest models on the training set, where the model including all individual HoNOS items and the age and sex of patients showed the best overall performance with an RMSE of 0.237 and an R-squared of 0.218. When applied to the test set, this resulted in an R-squared of 0.233 and a mean absolute error of 0.177.ConclusionsThe HoNOS can be mapped onto EQ-5D-3L utilities with moderate predictive accuracy. The reported mapping algorithm may be sufficient to predict overall population utility scores for use in health economic evaluations but lacks accuracy for individual patient predictions.

Identification of Risk Categories from the Advanced-Stage Hodgkin International Prognostic Index (A-HIPI) Model: A Detailed Analysis from the Hodgkin Lymphoma International Study for Individual Care (HoLISTIC) Consortium

Background: Prognostic modeling allows personalized risk prediction for individual patients (pt). The A-HIPI model in advanced stage classical Hodgkin lymphoma (AS-HL), developed and validated by the HoLISTIC Consortium (www.hodgkinconsortium.org), generates the individualized probability of a progression-free survival (PFS) event or death (OS) within the first 5 years (y) from diagnosis in pts based on continuous variables (www.qxmd.com/calculate/calculator_869/a-hipi). Clinical prognostic tools in lymphoma (eg, IPS, IPI, FLIPI, etc) typically use groupings of categorical values to define risk. Grouping a continuous value often results in loss of information, and most tools are not predictive for individual pts. However, discrete groupings have clinical utility &amp; practicality for a) defining pt populations for clinical trials &amp; real world studies, b) stratification within clinical trials, and c) crafting treatment guidelines. We studied potential approaches for utilizing the A-HIPI model to generate risk groups with input on strengths &amp; limitations from the HoLISTIC modeling team &amp; clinical experts. Methods: The A-HIPI model was developed via TRIPOD guidelines on 4,022 pts treated on 8 international clinical trials for AS-HL (Rodday. JCO 2023). External validation was performed on a dataset of 1,431 pts from 4 prospective registries. The 5y PFS (PFS5) in the A-HIPI development dataset was 77% (95% CI: 76-78); the 5y OS (OS5) was 92% (95% CI: 91-93). This represented the average outcome for a pt with AS-HL pt naïve to other clinical data at presentation. The distribution of PFS5 &amp; OS5 predictions were heavily skewed (ie, asymmetric distribution) in both the A-HIPI discovery and validation dataset. While not unexpected due to the excellent PFS &amp; OS in this disease setting, this presents challenges in the delineation of risk groups as depicted below . Three approaches were examined for the generation of A-HIPI risk groups. Proposed cutoffs were defined using the distribution of A-HIPI risk scores and data from the model-building cohort (ie, clinical trials). Validation was done using the A-HIPI validation cohort (ie, HL registries). Results: Approach 1: Risk groups based on clinical thresholds. Clinicians were queried what estimates of PFS5 would constitute high vs low risk. The positive, right-skewed distribution of A-HIPI risk scores limited this approach ( Figure), as cutoffs of PFS5 &amp;lt;70 and PFS5&amp;gt; 90 would only identify 15% and &amp;lt;1% of the population, respectively. Approach 2: Risk groups based on deviation from “average” pt. The 5-y PFS was 77% (95% CI: 76-78). We explored defining “standard risk” based on this confidence interval as well as clinical boundaries, with pts above or below this classified as decreased or increased risk, respectively. Results of a +/-5% clinical boundary are presented in the Table. Approach 3: Risk groups based on “ranking” of pts. Here we ranked the A-HIPI risk scores of the 4022 AS-HL pts in the model building cohort and used the distribution of the risk scores as a benchmark. The risk profile for a future pt is then compared to this distribution (eg, how do you rank compared to your peers). Continuous results are often presented this way (eg, tertiles or quartiles). It also allows flexibility for the user to define the size of the risk groups and/or clinical threshold of interest. Application of this approach showed good alignment between the predicted model percentiles and the observed distribution of scores in the validation cohort ( Table). This is also reflected in calibration curves presented in the primary manuscript. An online application (eg, R-Shiny) will be provided at the meeting to allow users to define their own cutoffs to aid in pt prognostication as well as identify populations for clinical trial development. Conclusions: There are challenges with defining risk groups from individual risk prediction modeling in AS-HL. Different applications and purposes, the skewed distribution of events/risk estimates, as well as varying clinical definitions of high risk, make it challenging to define consensus expert-based risk groupings for AS-HL. A flexible “rank-based” approach appeared to provide the most clinical utility &amp; data granularity, which may be leveraged for clinical trial design and pt stratification. Further analysis and discussion of how to optimally define high-risk and low-risk populations in AS-HL will be needed as new therapeutic options emerge.

End-to-end risk prediction of atrial fibrillation from the 12-Lead ECG by deep neural networks

BackgroundAtrial fibrillation (AF) is one of the most common cardiac arrhythmias that affects millions of people each year worldwide and it is closely linked to increased risk of cardiovas- cular diseases such as stroke and heart failure. Machine learning methods have shown promising results in evaluating the risk of developing atrial fibrillation from the electrocardiogram. We aim to develop and evaluate one such algorithm on a large CODE dataset collected in Brazil. MethodsWe used the CODE cohort to develop and test a model for AF risk prediction for individual patients from the raw ECG recordings without the use of additional digital biomarkers. The cohort is a collection of ECG recordings and annotations by the Telehealth Network of Minas Gerais, in Brazil. A convolutional neural network based on a residual network architecture was implemented to produce class probabilities for the classification of AF. The probabilities were used to develop a Cox proportional hazards model and a Kaplan-Meier model to carry out survival analysis. Hence, our model is able to perform risk prediction for the development of AF in patients without the condition. ResultsThe deep neural network model identified patients without indication of AF in the presented ECG but who will develop AF in the future with an AUC score of 0.845. From our survival model, we obtain that patients in the high-risk group (i.e. with the probability of a future AF case being >0.7) are 50% more likely to develop AF within 40 weeks, while patients belonging to the minimal-risk group (i.e. with the probability of a future AF case being less than or equal to 0.1) have >85% chance of remaining AF free up until after seven years. ConclusionWe developed and validated a model for AF risk prediction. If applied in clinical practice, the model possesses the potential of providing valuable and useful information in decision- making and patient management processes.

Precision medicine for traumatic cervical spinal cord injuries: accessible and interpretable machine learning models to predict individualized in-hospital outcomes

BACKGROUND CONTEXTA traumatic spinal cord injury (SCI) can cause temporary or permanent motor and sensory impairment, leading to serious short and long-term consequences that can result in significant morbidity and mortality. The cervical spine is the most commonly affected area, accounting for about 60% of all traumatic SCI cases. PURPOSEThis study aims to employ machine learning (ML) algorithms to predict various outcomes, such as in-hospital mortality, nonhome discharges, extended length of stay (LOS), extended length of intensive care unit stay (ICU-LOS), and major complications in patients diagnosed with cervical SCI (cSCI). STUDY DESIGNOur study was a retrospective machine learning classification study aiming to predict the outcomes of interest, which were binary categorical variables, in patients diagnosed with cSCI. PATIENT SAMPLEThe data for this study were obtained from the American College of Surgeons (ACS) Trauma Quality Program (TQP) database, which was queried to identify patients who suffered from cSCI between 2019 and 2021. OUTCOME MEASURESThe outcomes of interest of our study were in-hospital mortality, nonhome discharges, prolonged LOS, prolonged ICU-LOS, and major complications. The study evaluated the models' performance using both graphical and numerical methods. The receiver operating characteristic (ROC) and precision-recall curves (PRC) were used to assess model performance graphically. Numerical evaluation metrics included AUROC, balanced accuracy, weighted area under PRC (AUPRC), weighted precision, and weighted recall. METHODSThe study employed data from the American College of Surgeons (ACS) Trauma Quality Program (TQP) database to identify patients with cSCI. Four ML algorithms, namely XGBoost, LightGBM, CatBoost, and Random Forest, were utilized to develop predictive models. The most effective models were then incorporated into a publicly available web application designed to forecast the outcomes of interest. RESULTSThere were 71,661 patients included in the analysis for the outcome mortality, 67,331 for the outcome nonhome discharges, 76,782 for the outcome prolonged LOS, 26,615 for the outcome prolonged ICU-LOS, and 72,132 for the outcome major complications. The algorithms exhibited an AUROC value range of 0.78 to 0.839 for in-hospital mortality, 0.806 to 0.815 for nonhome discharges, 0.679 to 0.742 for prolonged LOS, 0.666 to 0.682 for prolonged ICU-LOS, and 0.637 to 0.704 for major complications. An open access web application was developed as part of the study, which can generate predictions for individual patients based on their characteristics. CONCLUSIONSOur study suggests that ML models can be valuable in assessing risk for patients with cervical cSCI and may have considerable potential for predicting outcomes during hospitalization. ML models demonstrated good predictive ability for in-hospital mortality and nonhome discharges, fair predictive ability for prolonged LOS, but poor predictive ability for prolonged ICU-LOS and major complications. Along with these promising results, the development of a user-friendly web application that facilitates the integration of these models into clinical practice is a significant contribution of this study. The product of this study may have significant implications in clinical settings to personalize care, anticipate outcomes, facilitate shared decision making and informed consent processes for cSCI patients.

Interpretable machine learning models to predict short-term postoperative outcomes following posterior cervical fusion.

By predicting short-term postoperative outcomes before surgery, patients who undergo posterior cervical fusion (PCF) surgery may benefit from more precise patient care plans that reduce the likelihood of unfavorable outcomes. We developed machine learning models for predicting short-term postoperative outcomes and incorporate these models into an open-source web application in this study. The American College of Surgeons National Surgical Quality Improvement Program database was used to identify patients who underwent PCF surgery. Prolonged length of stay, non-home discharges, and readmissions were the three outcomes that were investigated. To predict these three outcomes, machine learning models were developed and incorporated into an open access web application. A total of 6277 patients that underwent PCF surgery were included in the analysis. The most accurately predicted outcome in terms of the area under the receiver operating characteristic curve (AUROC) was the non-home discharges with a mean AUROC of 0.812, and the most accurately predicting algorithm in terms of AUROC was the LightGBM algorithm with a mean AUROC of 0.766. The following URL will take users to the open access web application created to provide predictions for individual patients based on their characteristics: https://huggingface.co/spaces/MSHS-Neurosurgery-Research/NSQIP-PCF. Machine learning techniques have a significant potential for predicting postoperative outcomes following PCF surgery. The development of predictive models as clinically useful decision-making tools may significantly improve risk assessment and prognosis as the amount of data in spinal surgery keeps growing. Here, we present predictive models for PCF surgery that are meant to accomplish the aforementioned goals and make them publicly available.

Personalized event prediction for Electronic Health Records

Clinical event sequences consist of hundreds of clinical events that represent records of patient care in time. Developing accurate predictive models of such sequences is of a great importance for supporting a variety of models for interpreting/classifying the current patient condition, or predicting adverse clinical events and outcomes, all aimed to improve patient care. One important challenge of learning predictive models of clinical sequences is their patient-specific variability. Based on underlying clinical conditions, each patient’s sequence may consist of different sets of clinical events (observations, lab results, medications, procedures). Hence, simple population-wide models learned from event sequences for many different patients may not accurately predict patient-specific dynamics of event sequences and their differences. To address the problem, we propose and investigate multiple new event sequence prediction models and methods that let us better adjust the prediction for individual patients and their specific conditions. The methods developed in this work pursue refinement of population-wide models to subpopulations, self-adaptation, and a meta-level model switching that is able to adaptively select the model with the best chance to support the immediate prediction. We analyze and test the performance of these models on clinical event sequences of patients in MIMIC-III database.

Prognosis at Your Fingertips: A Machine Learning-Based Web Application for Outcome Prediction in Acute Traumatic Epidural Hematoma.

Traumatic brain injury (TBI) affects 69 million people worldwide each year, and acute traumatic epidural hematoma (atEDH) is a frequent and severe consequence of TBI. The aim of the study is to use machine learning (ML) algorithms to predict in-hospital death, non-home discharges, prolonged length of stay (LOS), prolonged length of intensive care unit stay (ICU-LOS), and major complications in patients with atEDH and incorporate the resulting ML models into a user-friendly web application for use in the clinical settings. The American College of Surgeons (ACS) Trauma Quality Program (TQP) database was used to identify patients with atEDH. Four ML algorithms (XGBoost, LightGBM, CatBoost, and Random Forest) were utilized, and the best performing models were incorporated into an open-access web application to predict the outcomes of interest. The study found that the ML algorithms had high area under the receiver operating characteristic curve (AUROC) values in predicting outcomes for patients with atEDH. In particular, the algorithms had an AUROC value range of between 0.874 to 0.956 for in-hospital mortality, 0.776 to 0.798 for non-home discharges, 0.737 to 0.758 for prolonged LOS, 0.712 to 0.774 for prolonged ICU-LOS, and 0.674 to 0.733 for major complications. The following link will take users to the open-access web application designed to generate predictions for individual patients based on their characteristics: huggingface.co/spaces/MSHS-Neurosurgery-Research/TQP-atEDH. This study aimed to improve the prognostication of patients with atEDH using ML algorithms and developed a web application for easy integration in clinical practice. It found that ML algorithms can aid in risk stratification and have significant potential for predicting in-hospital outcomes. Results demonstrated excellent performance for predicting in-hospital death and fair performance for non-home discharges, prolonged LOS and ICU-LOS, and poor performance for major complications.

Machine Learning-Based Prediction of Short-Term Adverse Postoperative Outcomes in Cervical Disc Arthroplasty Patients

This study aimed to assess the effectiveness of machine learning (ML) algorithms in predicting short-term adverse postoperative outcomes after cervical disc arthroplasty (CDA) and to create a user-friendly and accessible tool for this purpose. The American College of Surgeons National Surgical Quality Improvement Program database was used to identify patients who underwent CDA. The outcome of interest was the combined occurrence of adverse events in the short-term postoperative period, including prolonged stay, major complications, non-home discharges, and 30-day readmissions. To predict the combined outcome of interest, short-term adverse postoperative outcomes, four different ML algorithms were utilized to develop predictive models, and these models were incorporated into an open access web application. A total of 6604 patients that underwent CDA were included in the analysis. The mean area under the receiver operating characteristic curve (AUROC) and accuracy were 0.814 and 87.8% for all algorithms. SHapley Additive exPlanations (SHAP) analyses revealed that white race was the most important predictor variable for all four algorithms. The following URL will take users to the open access web application created to provide predictions for individual patients based on their characteristics: huggingface.co/spaces/MSHS-Neurosurgery-Research/NSQIP-CDA. ML approaches have the potential to predict postoperative outcomes after CDA surgery. As the amount of data in spinal surgery grows, the development of predictive models as clinically useful decision-making tools may significantly improve risk assessment and prognosis. We present and make publicly available predictive models for CDA intended to achieve the goals mentioned above.

A Machine Learning-Based Online Prediction Tool for Predicting Short-Term Postoperative Outcomes Following Spinal Tumor Resections.

Background: Preoperative prediction of short-term postoperative outcomes in spinal tumor patients can lead to more precise patient care plans that reduce the likelihood of negative outcomes. With this study, we aimed to develop machine learning algorithms for predicting short-term postoperative outcomes and implement these models in an open-source web application. Methods: Patients who underwent surgical resection of spinal tumors were identified using the American College of Surgeons, National Surgical Quality Improvement Program. Three outcomes were predicted: prolonged length of stay (LOS), nonhome discharges, and major complications. Four machine learning algorithms were developed and integrated into an open access web application to predict these outcomes. Results: A total of 3073 patients that underwent spinal tumor resection were included in the analysis. The most accurately predicted outcomes in terms of the area under the receiver operating characteristic curve (AUROC) was the prolonged LOS with a mean AUROC of 0.745 The most accurately predicting algorithm in terms of AUROC was random forest, with a mean AUROC of 0.743. An open access web application was developed for getting predictions for individual patients based on their characteristics and this web application can be accessed here: huggingface.co/spaces/MSHS-Neurosurgery-Research/NSQIP-ST. Conclusion: Machine learning approaches carry significant potential for the purpose of predicting postoperative outcomes following spinal tumor resections. Development of predictive models as clinically useful decision-making tools may considerably enhance risk assessment and prognosis as the amount of data in spinal tumor surgery continues to rise.

Optimal mode of delivery in pregnancy: Individualized predictions using national vital statistics data.

Child birth via Cesarean section accounts for approximately 32% of all births each year in the United States. A variety of risk factors and complications can lead caregivers and patients to plan for a Cesarean delivery in advance before onset of labor. However, a non-trivial subset of Cesarean sections (∼25%) are unplanned and occur after an initial trial of labor is attempted. Unfortunately, patients who deliver via unplanned Cesarean sections have increased maternal morbidity and mortality rates and higher rates of neonatal intensive care admissions. In an effort to develop models aimed at improving health outcomes in labor and delivery, this work seeks to explore the use of national vital statistics data to quantify the likelihood of an unplanned Cesarean section based on 22 maternal characteristics. Machine learning techniques are used to ascertain influential features, train and evaluate models, and assess accuracy against available test data. Based on cross-validation results from a large training cohort (n = 6,530,467 births), the gradient-boosted tree algorithm was identified as the best performer and was evaluated on a large test cohort (n = 10,613,877 births) for two prediction scenarios. Area under the receiver operating characteristic curves of 0.77 or higher and recall scores of 0.78 or higher were obtained and the resulting models are well calibrated. Combined with feature importance analysis to explain why certain maternal characteristics lead to a specific prediction in individual patients, the developed analysis pipeline provides additional quantitative information to aid in the decision process on whether to plan for a Cesarean section in advance, a substantially safer option among women at a high risk of unplanned Cesarean delivery during labor.

Incorporating progesterone receptor expression into the PREDICT breast prognostic model

BackgroundPredict Breast (www.predict.nhs.uk) is an online prognostication and treatment benefit tool for early invasive breast cancer. The aim of this study was to incorporate the prognostic effect of progesterone receptor (PR) status into a new version of PREDICT and to compare its performance to the current version (2.2). MethodThe prognostic effect of PR status was based on the analysis of data from 45,088 European patients with breast cancer from 49 studies in the Breast Cancer Association Consortium. Cox proportional hazard models were used to estimate the hazard ratio for PR status. Data from a New Zealand study of 11,365 patients with early invasive breast cancer were used for external validation. Model calibration and discrimination were used to test the model performance. ResultsHaving a PR-positive tumour was associated with a 23% and 28% lower risk of dying from breast cancer for women with oestrogen receptor (ER)-negative and ER-positive breast cancer, respectively. The area under the ROC curve increased with the addition of PR status from 0.807 to 0.809 for patients with ER-negative tumours (p = 0.023) and from 0.898 to 0.902 for patients with ER-positive tumours (p = 2.3 × 10−6) in the New Zealand cohort. Model calibration was modest with 940 observed deaths compared to 1151 predicted. ConclusionThe inclusion of the prognostic effect of PR status to PREDICT Breast has led to an improvement of model performance and more accurate absolute treatment benefit predictions for individual patients. Further studies should determine whether the baseline hazard function requires recalibration.