Patient-specific Predictive Model Research Articles

Addressing data limitations in seizure prediction through transfer learning

According to the literature, seizure prediction models should be developed following a patient-specific approach. However, seizures are usually very rare events, meaning the number of events that may be used to optimise seizure prediction approaches is limited. To overcome such constraint, we analysed the possibility of using data from patients from an external database to improve patient-specific seizure prediction models. We present seizure prediction models trained using a transfer learning procedure. We trained a deep convolutional autoencoder using electroencephalogram data from 41 patients collected from the EPILEPSIAE database. Then, a bidirectional long short-term memory and a classifier layers were added on the top of the encoder part and were optimised for 24 patients from the Universitätsklinikum Freiburg individually. The encoder was used as a feature extraction module. Therefore, its weights were not changed during the patient-specific training. Experimental results showed that seizure prediction models optimised using pretrained weights present about four times fewer false alarms while maintaining the same ability to predict seizures and achieved more 13% validated patients. Therefore, results evidenced that the optimisation using transfer learning was more stable and faster, saving computational resources. In summary, adopting transfer learning for seizure prediction models represents a significant advancement. It addresses the data limitation seen in the seizure prediction field and offers more efficient and stable training, conserving computational resources. Additionally, despite the compact size, transfer learning allows to easily share data knowledge due to fewer ethical restrictions and lower storage requirements. The convolutional autoencoder developed in this study will be shared with the scientific community, promoting further research.

Same therapy, same calcium mobilization? Exploring calcium exchange across body compartments using a patient-specific predictive model.

Comprehensive, patient-specific models are essential to study calcium deposition and mobilization during dialysis. We aim to develop tools to support clinical prescriptions with a more accurate approach for the prediction of calcium mobilization while also considering major electrolytes and catabolites. We modified a multi-solute model predicting patient-specific dialysis response by incorporating a calcium buffer to represent bone exchanges. Data from four centers, involving 127 patients with six sessions each, were utilized. For each patient, three sessions were allocated for model training (ID123), while the remaining sessions were for validation (PRED456). The normalized root mean square error (nRMSE%) was used to evaluate both descriptive and predictive accuracy. Correlations between initial data and calcium exchanges were also assessed. The overall nRMSE% for ID123 was 3.92%. For PRED456, it was 3.46% (ranging from a minimum of 1.17% for [Na+] to a maximum of 6.62% for [urea]). The median nRMSE% for plasma calcium varied between 1.13 and 8.32 for SHD sessions, depending on whether Ca_dialysis fluid (Cad) was ≥ or <1.50 mmol/L, respectively. For HDF sessions, the range was between 2.90 and 5.89. A significant and moderate correlation was found between overall calcium removal and the buffer balance. The most robust correlation observed was between the amount of calcium administered via post-dilution fluid and the overall calcium removal in the dialysis filter. Identical therapy settings do not uniformly affect calcium mobilization, and our approach offers insight into calcium distribution across body compartments. This understanding will enhance clinical prescription practices.

#1944 How the patient characteristics and the prescription can be accounted for in patient-specific predictive HD models?

Abstract Background and Aims Compartmental models in HemoDialysis (HD) allow for understanding and predicting the patient-specific response to the treatment. These models, after training with consistent clinical data, can be adapted to each single patient through the estimation of a set of parameters. This work explores the relationship between the sets of estimated parameters and the anamneses of the enrolled patients, together with the settings of the HD sessions, and their time evolution. The finding of similar relationships will help properly set the models even avoiding the training phase. Method The data acquired during the INTERREG InterACTIVE-HD 2.0 observational longitudinal study, in partnership with 5 Italian and Switzerland hospitals (ASSTs Lariana, SetteLaghi, Valtellina e Alto Lario, Regional Hospital of Lugano, and Kantonsspital Graubünden of Chur Dialysis Units), were used (576 sessions). Clinical prescription, machine settings, and blood composition at the start, end, and every hour during the HDs were monitored over 7 months. The parametric multi-compartmental, multi-solute model optimized during the study was used. The patient-specific parameters are: Their default values (literature-based) are: Patient-specific parameters have been identified for each patient, by training the model with 3 session's data. Then the model can be used to predict the patient's response by knowing the clinical data at the beginning of the HD and setting the parameters equal to the mean of the 3 identified values. A correlation analysis between the patient-dependent parameters and the anamneses of the enrolled patients, together with the settings of the HD sessions, has been performed. The optimization orders have been also considered. Results k(K+), k(Urea), and k(Crea) are the first optimized and remain generally of the same magnitude. k(Na+), k(Cl-), k(HCO3-), k(Mg2+), and k(Glu), correlate with the UF rate, the therapy technique, or the concentrations in dialysate bags. They are later optimized and characterized by wider variability in values and optimization orders. k(Ca2+) remains at the default in 98% of the sessions, being optimized last. More convective therapies seem to induce patient-specific variations in active and passive cellular transport for specific solutes. η relevant correlations refer to the dialysate bag composition and the baseline conditions of the patients. They are in general late optimized, in HDF before than in SHD. η(Urea) and η(Crea) in 99% of the cases do not move from the default, suggesting a lower interaction of the non-ionic species with the filter membrane. They are optimized only in non-hypertensive insulin-dependent diabetic patients. No correlation between η and the filter material or the initial serum albumin was instead observed. ρ is related to the therapy; in HDF sessions it is over the range boundaries ([0,10]). For females undergoing SHD, it assumes the highest values (&amp;gt;9). ρ seems also to be affected by the use of beta-blockers (p &amp;lt; 0.01) and antihyperglycemic ((p &amp;lt; 0.01). Ca_buff coefficients for females (higher values) and males are statistically different. Diabetics using anti-hyperglycemia drugs are characterized by lower values. Conclusion Compartmental patient-specific models, if successfully trained become a useful support tool for therapy customization, mainly for ESRD patients with multiple comorbidities. The observed correlations represent the first step to causally relate the values of the modulating parameters with the patient characteristics implying the reduction or amplification of a specific response to the treatment. The next step will be an analysis of combined effects. The final goal is to be able to correctly set up the models even before the training phase is completed.

Abstract 907: ScreenDL: A transfer learning framework integrating tumor omics and functional drug screening for personalized clinical drug response prediction

Abstract Precision oncology hinges on accurate prediction of patient-specific treatment response from tumor molecular inputs. While deep learning (DL) models achieve state-of-the-art response prediction in cell lines, existing methods do not readily translate to the clinic where training data is limited. Here, we have developed and validated ScreenDL, a novel DL framework designed explicitly for use in clinical precision oncology applications. The underlying architecture of ScreenDL consists of two fully connected branches dedicated to learning rich embeddings of a drug’s chemical structure and a tumor’s transcriptomic profile respectively. These tumor omic and drug molecular embeddings are fed into a shared response prediction subnetwork which outputs predicted ln(IC50) values. The ScreenDL training schema incorporates three phases, each designed to improve clinical response prediction: (1) an initial pretraining phase leverages large-scale cell line pharmacogenomic datasets to extract patterns/relationships linking tumor omics and drug response; (2) subsequent transfer learning integrates pharmacogenomic data from patient-derived models of cancer, adapting the pretrained model to a more patient-relevant context; (3) patient-level omics and drug screening data (e.g., from matched patient-derived organoid (PDO) models) is integrated through patient-specific fine-tuning, generating a patient-specific response prediction model. Importantly, biomaterial from surgical tumor resection is often available for organoid establishment and patient-specific drug screening in practical clinical scenarios. To assess the utility of our ScreenDL framework for clinical response prediction, we applied ScreenDL to predict treatment response in 50 advanced/metastatic triple negative breast cancer (TNBC) patient-derived xenograft models (PDX). We leveraged tumor omics profiles for each PDX and drug screening data from matched PDX-derived organoids (PDxOs), mirroring the combination of patient-level tumor omic characterization and functional drug screening in matched PDOs, a protocol currently being piloted in Functional Precision Oncology trials at the University of Utah Huntsman Cancer Institute. After cell line pretraining, ScreenDL achieves a modest median Pearson correlation per drug of 0.15 relative to 0.03 for the leading compared model. However, transfer learning and subsequent PDX-specific fine-tuning significantly improve performance, producing median Pearson correlations per drug of 0.39 and 0.51, respectively. These results demonstrate the ability of our ScreenDL framework to dramatically improve response prediction in clinically relevant domains, bringing DL-based precision treatment selection closer to clinical application. Citation Format: Casey Sederman, Tonya Di Sera, Yi Qiao, Xiaomeng Huang, Bryan E. Welm, Alana L. Welm, Gabor Marth. ScreenDL: A transfer learning framework integrating tumor omics and functional drug screening for personalized clinical drug response prediction [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 907.

How microsimulation translates outcome estimates to patient lifetime event occurrence in the setting of heart valve disease.

Treatment decisions in healthcare often carry lifelong consequences that can be challenging to foresee. As such, tools that visualize and estimate outcome after different lifetime treatment strategies are lacking and urgently needed to support clinical decision-making in the setting of rapidly evolving healthcare systems, with increasingly numerous potential treatments. In this regard, microsimulation models may prove valuable additions to current risk-prediction models. Notable advantages of microsimulation encompass input from multiple data sources, the ability to move beyond time-to-first-event analysis, accounting for multiple types of events, and generating projections of lifelong outcomes. This review aims to clarify the concept of microsimulation, also known as individualized state-transition models, and help clinicians better understand its potential in clinical decision-making. A practical example of a patient with heart valve disease is used to illustrate key components of microsimulation models, such as health states, transition probabilities, input parameters (e.g., evidence-based risks of events), and various aspects of mortality. Finally, this review focuses on future efforts needed in microsimulation to allow for increasing patient-tailoring of the models by extending the general structure with patient-specific prediction models and translating them to meaningful, user-friendly tools that may be used by both clinician and patient to support clinical decision-making.

Longitudinal prognosis of Parkinson’s outcomes using causal connectivity

Despite the prevalence of Parkinson’s disease (PD), there are no clinically-accepted neuroimaging biomarkers to predict the trajectory of motor or cognitive decline or differentiate Parkinson’s disease from atypical progressive parkinsonian diseases. Since abnormal connectivity in the motor circuit and basal ganglia have been previously shown as early markers of neurodegeneration, we hypothesize that patterns of interregional connectivity could be useful to form patient-specific predictive models of disease state and of PD progression. We use fMRI data from subjects with Multiple System Atrophy (MSA), Progressive Supranuclear Palsy (PSP), idiopathic PD, and healthy controls to construct predictive models for motor and cognitive decline and differentiate between the four subgroups. Further, we identify the specific connections most informative for progression and diagnosis. When predicting the one-year progression in the MDS-UPDRS-III1*1*A summary of all acronyms used in this paper is presented in Appendix A. Supplementary Data. and Montreal Cognitive assessment (MoCA), we achieve new state-of-the-art mean absolute error performance. Additionally, the balanced accuracy we achieve in the diagnosis of PD, MSA, PSP, versus healthy controls surpasses that attained in most clinics, underscoring the relevance of the brain connectivity features. Our models reveal the connectivity between deep nuclei, motor regions, and the thalamus as the most important for prediction. Collectively these results demonstrate the potential of fMRI connectivity as a prognostic biomarker for PD and increase our understanding of this disease.

Prognostic nomogram for predicting lower extremity deep venous thrombosis in ruptured intracranial aneurysm patients who underwent endovascular treatment.

Lower extremity deep vein thrombosis (DVT) is one of the major postoperative complications in patients with ruptured intracranial aneurysms (RIA) who underwent endovascular treatment (EVT). However, patient-specific predictive models are still lacking. This study aimed to construct and validate a nomogram model for estimating the risk of lower extremity DVT for RIA patients who underwent EVT. This cohort study enrolled 471 RIA patients who received EVT in our institution between 1 January 2020 to 4 February 2022. Perioperative information on participants is collected to develop and validate a nomogram for predicting lower extremity DVT in RIA patients after EVT. Predictive accuracy, discriminatory capability, and clinical effectiveness were evaluated by concordance index (C-index), calibration curves, and decision curve analysis. Multivariate logistic regression analysis showed that age, albumin, D-dimer, GCS score, middle cerebral artery aneurysm, and delayed cerebral ischemia were independent predictors for lower extremity DVT. The nomogram for assessing individual risk of lower extremity DVT indicated good predictive accuracy in the primary cohort (c-index, 0.92) and the validation cohort (c-index, 0.85), with a wide threshold probability range (4-82%) and superior net benefit. The present study provided a reliable and convenient nomogram model developed with six optimal predictors to assess postoperative lower extremity DVT in RIA patients, which may benefit to strengthen the awareness of lower extremity DVT control and supply appropriate resources to forecast patients at high risk of RIA-related lower extremity DVT.

Evaluating motion of pancreatic tumors and anatomical surrogates using cine MRI in 0.35T MRgRT under free breathing conditions.

Treatment tolerability is a significant limitation to pancreatic cancer treatment with radiotherapy due to proximity to highly radiosensitive organs and respiratory motion necessitating expanded target margins. Further, pancreatic tumors are difficult to visualize on conventional radiotherapy systems. Surrogates are often used to locate the tumor but are often inconsistent and do not provide strong positional relations throughout the respiratory cycle. This work utilizes a retrospective dataset of 45 pancreatic cancer patients treated on an MR-Linac system with cine MRI acquired for real-time target tracking. We investigated intra-fraction motion of tumors and two abdominal surrogates, leading to prediction models between the tumor and surrogate. Patient specific motion evaluation and prediction models were generated from 225 cine MRI series acquired during treatment. Tumor contours were used to evaluate the pancreatic tumor motion. Linear regression and principal component analysis (PCA) based models were used to predict tumor position from the anterior-posterior (AP) motion of the abdominal surface, the superior-inferior (SI) motion of the diaphragm, or a combination. Models were evaluated using mean squared error (MSE) and mean absolute error (MAE). Contour analysis showed the average pancreatic tumor motion range was 7.4±2.7mm and 14.9±5.8mm in the AP and SI directions, respectively. The PCA model had MSE of 1.4mm2 and 0.6mm2 , for the SI and AP directions, respectively, with both surrogates as inputs for the models. When only the abdomen surrogate was used, MSE was 1.3mm2 and 0.4mm2 in the SI and AP directions, while it was 0.4mm2 and 1.3mm2 when only the diaphragm surrogate was used. We evaluated intra-fraction pancreatic tumor motion and demonstrated prediction models between the tumor and surrogate. The models calculated the pancreatic tumor position from diaphragm, abdominal, or both contours within standard pancreatic cancer target margin, and the process could be applied to other disease sites in the abdominothoracic cavity.

Removing artefacts and periodically retraining improve performance of neural network-based seizure prediction models

The development of seizure prediction models is often based on long-term scalp electroencephalograms (EEGs) since they capture brain electrical activity, are non-invasive, and come at a relatively low-cost. However, they suffer from major shortcomings. First, long-term EEG is usually highly contaminated with artefacts. Second, changes in the EEG signal over long intervals, known as concept drift, are often neglected. We evaluate the influence of these problems on deep neural networks using EEG time series and on shallow neural networks using widely-used EEG features. Our patient-specific prediction models were tested in 1577 hours of continuous EEG, containing 91 seizures from 41 patients with temporal lobe epilepsy who were undergoing pre-surgical monitoring. Our results showed that cleaning EEG data, using a previously developed artefact removal method based on deep convolutional neural networks, improved prediction performance. We also found that retraining the models over time reduced false predictions. Furthermore, the results show that although deep neural networks processing EEG time series are less susceptible to false alarms, they may need more data to surpass feature-based methods. These findings highlight the importance of robust data denoising and periodic adaptation of seizure prediction models.

Correlation between Collagen Type I/III Ratio and Scar Formation in Patients Undergoing Immediate Reconstruction with the Round Block Technique after Breast-Conserving Surgery.

The aim of this study was to investigate the relationship between the collagen type I/III ratio and scarring in patients who underwent immediate reconstruction with the round block technique (RBT) after breast conservation surgery. Seventy-eight patients were included, and demographic and clinical characteristics were recorded. The collagen type I/III ratio was measured using immunofluorescence staining and digital imaging, and scarring was assessed using the Vancouver Scar Scale (VSS). The mean VSS scores were 1.92 ± 2.01 and 1.79 ± 1.89, as assessed by two independent plastic surgeons, with good reliability of the scores. A significant positive correlation was found between VSS and the collagen type I/III ratio (r = 0.552, p < 0.01), and a significant negative correlation was found between VSS and the collagen type III content (r = -0.326, p < 0.05). Multiple linear regression analysis showed that the collagen type I/III ratio had a significant positive effect on VSS (β = 0.415, p = 0.028), whereas the collagen type I and collagen type III content had no significant effect on VSS. These findings suggest that the collagen type I/III ratio is associated with scar development in patients undergoing RBT after breast conservation surgery. Further research is needed to develop a patient-specific scar prediction model based on genetic factors affecting the collagen type I/III ratio.

Cross-Domain Transfer of EEG to EEG or ECG Learning for CNN Classification Models.

Electroencephalography (EEG) is often used to evaluate several types of neurological brain disorders because of its noninvasive and high temporal resolution. In contrast to electrocardiography (ECG), EEG can be uncomfortable and inconvenient for patients. Moreover, deep-learning techniques require a large dataset and a long time for training from scratch. Therefore, in this study, EEG-EEG or EEG-ECG transfer learning strategies were applied to explore their effectiveness for the training of simple cross-domain convolutional neural networks (CNNs) used in seizure prediction and sleep staging systems, respectively. The seizure model detected interictal and preictal periods, whereas the sleep staging model classified signals into five stages. The patient-specific seizure prediction model with six frozen layers achieved 100% accuracy for seven out of nine patients and required only 40 s of training time for personalization. Moreover, the cross-signal transfer learning EEG-ECG model for sleep staging achieved an accuracy approximately 2.5% higher than that of the ECG model; additionally, the training time was reduced by >50%. In summary, transfer learning from an EEG model to produce personalized models for a more convenient signal can both reduce the training time and increase the accuracy; moreover, challenges such as data insufficiency, variability, and inefficiency can be effectively overcome.

A survival prediction model via interpretable machine learning for patients with oropharyngeal cancer following radiotherapy.

To explore interpretable machine learning (ML) methods, with the hope of adding more prognosis value, for predicting survival for patients with Oropharyngeal-Cancer (OPC). A cohort of 427 OPC patients (Training 341, Test 86) from TCIA databasewas analyzed. Radiomic features of gross-tumor-volume (GTV) extracted from planning CT using Pyradiomics, and HPV p16 status, etc. patient characteristics were considered as potential predictors. A multi-level dimension reduction algorithm consisting of Least-Absolute-Selection-Operator (Lasso) and Sequential-Floating-Backward-Selection (SFBS) was proposed to effectively remove redundant/irrelevant features. The interpretable model was constructed by quantifying the contribution of each feature to the Extreme-Gradient-Boosting (XGBoost) decision by Shapley-Additive-exPlanations (SHAP) algorithm. The Lasso-SFBS algorithm proposed in this study finally selected 14 features, and our prediction model achieved an area-under-ROC-curve (AUC) of 0.85 on the test dataset based on this feature set. The ranking of the contribution values calculated by SHAP shows that the top predictors that were most correlated with survival were ECOG performance status, wavelet-LLH_firstorder_Mean, chemotherapy, wavelet-LHL_glcm_InverseVariance, tumor size. Those patients who had chemotherapy, with positive HPV p16 status, and lower ECOG performance status, tended to have higher SHAP scores and longer survival; who had an older age at diagnosis, heavy drinking and smoking pack year history, tended to lower SHAP scores and shorter survival. We demonstrated predictive values of combined patient characteristics and imaging features for the overall survival of OPC patients. The multi-level dimension reduction algorithm can reliably identify the most plausible predictors that are mostly associated with overall survival. The interpretable patient-specific survival prediction model, capturing correlations of each predictor and clinical outcome, was developed to facilitate clinical decision-making for personalized treatment.

Predictive Model of Liver Toxicity to Aid the Personalized Selection of Proton Versus Photon Therapy in Hepatocellular Carcinoma

Our objective was to develop an externally validated model for predicting liver toxicity after radiation therapy in patients with hepatocellular carcinoma (HCC) that can integrate both photon and proton dose distributions with patient-specific characteristics. Training data consisted of all patients with HCC treated between 2008 and 2019 at our institution (n=117, 60%/40% photon/proton). We developed a shallow convolutional neural network (CNN) to predict posttreatment liver dysfunction from the differential dose-volume histogram (DVH) and baseline liver metrics. To reduce bias and improve robustness, we used ensemble learning (CNNE). After a preregistered study analysis plan, we evaluated stability using internal bootstrap resampling and generalizability using a data set from a different institution (n=88). Finally, we implemented a class activation map method to characterize the critical DVH subregions and benchmarked the model against logistic regression and XGBoost. The models were evaluated using the area under the receiver operating characteristic curve and area under the precision-recall curve. The CNNE model showed similar internal performance and robustness compared with the benchmarks. CNNE exceeded the benchmark models in external validation, with an area under the receiver operating characteristic curve of 0.78 versus 0.55 to 0.70, and an area under the precision-recall curve of 0.6 versus 0.43 to 0.52. The model showed improved predictive power in the photon group, excellent specificity in both modalities, and high sensitivity in the photon high-risk group. Models built solely on DVHs confirm outperformance of the CNNE and indicate that the proposed structure efficiently abstracts features from both proton and photon dose distributions. The activation map method demonstrates the importance of the low-dose bath and its interaction with low liver function at baseline. We developed and externally validated a patient-specific prediction model for hepatic toxicity based on the entire DVH and clinical factors that can integrate both photon and proton therapy cohorts. This model complements the new American Society for Radiation Oncology clinical practice guidelines and could support value-driven integration of proton therapy into the management of HCC.

AMPREDICT PROsthetics—Predicting Prosthesis Mobility to Aid in Prosthetic Prescription and Rehabilitation Planning

ObjectiveTo develop and validate a patient-specific multivariable prediction model that uses variables readily available in the electronic medical record to predict 12-month mobility at the time of initial post-amputation prosthetic prescription. The prediction model is designed for patients who have undergone their initial transtibial (TT) or transfemoral (TF) amputation because of complications of diabetes and/or peripheral artery disease. DesignMulti-methodology cohort study that identified patients retrospectively through a large Veteran's Affairs (VA) dataset then prospectively collected their patient-reported mobility. SettingThe VA Corporate Data Warehouse, the National Prosthetics Patient Database, participant mailings, and phone calls. ParticipantsThree-hundred fifty-seven veterans who underwent an incident dysvascular TT or TF amputation and received a qualifying lower limb prosthesis between March 1, 2018, and November 30, 2020 (N=357). InterventionsNot applicable. Main Outcome MeasureThe Amputee Single Item Mobility Measure (AMPSIMM) was divided into a 4-category outcome to predict wheelchair mobility (0-2), and household (3), basic community (4), or advanced community ambulation (5-6). ResultsMultinomial logistic lasso regression, a machine learning methodology designed to select variables that most contribute to prediction while controlling for overfitting, led to a final model including 23 predictors of the 4-category AMPSIMM outcome that effectively discriminates household ambulation from basic community ambulation and from advanced community ambulation—levels of key clinical importance when estimating future prosthetic demands. The overall model performance was modest as it did not discriminate wheelchair from household mobility as effectively. ConclusionsThe AMPREDICT PROsthetics model can assist providers in estimating individual patients’ future mobility at the time of prosthetic prescription, thereby aiding in the formulation of appropriate mobility goals, as well as facilitating the prescription of a prosthetic device that is most appropriate for anticipated functional goals.

AMPREDICT PROsthetics- Predicting Prosthetic Mobility to Aid in Prosthetic Prescription and Rehabilitation Planning

Research Objectives To develop and validate a patient-specific multivariable prediction model that utilizes readily available predictors to predict 12-month mobility at the time of initial post amputation prosthetic prescription. The prediction model is designed to be used on patients who have undergone their initial transtibial (TT) or transfemoral (TF) amputation for chronic limb threatening ischemia (CLTI). The overall goal is to enhance prediction of future mobility to assist the rehabilitation team in setting appropriate, evidence-based goals, to better inform patient expectations, and to facilitate appropriate prosthetic prescription. Design Mixed-methods study that identified patients retrospectively through a large Veteran's Affairs (VA) dataset then prospectively collected patient reported mobility. Setting The Veterans Administration Corporate Data Warehouse (CDW), the National Prosthetics Patient Database (NPPD), participant mailings and phone calls. Participants 357 Veterans who underwent an incident TT or TF amputation for CLTI and received a qualifying lower limb prosthesis (LLP) between March 1, 2018, and November 30, 2020. Interventions Not applicable. Main Outcome Measures The Amputee Single Item Mobility Measure (AMPSIMM) was divided into a 4-category outcome to predict wheelchair mobility (0-2), and household (3), basic community (4), or advanced community ambulation (5-6). Results Variable selection, using a machine learning methodology, led to a final model of 23 predictors that effectively discriminates household ambulation from basic community ambulation and from advanced community ambulation – levels of key clinical importance when estimating future prosthetic demands. The overall model performance was modest as it did not discriminate wheelchair from household mobility as effectively. Conclusions The AMPREDICT PROsthetics model can assist providers in estimating individual patient's future mobility at the time of prosthetic prescription thereby aiding in the formulation of appropriate mobility goals, as well as facilitating the prescription of a prosthetic device that is most appropriate for their anticipated functional goals. Author(s) Disclosures No conflicts of interest. This project was completed from funding provided by a Veterans Administration Research Rehabilitation and Development Grant. To develop and validate a patient-specific multivariable prediction model that utilizes readily available predictors to predict 12-month mobility at the time of initial post amputation prosthetic prescription. The prediction model is designed to be used on patients who have undergone their initial transtibial (TT) or transfemoral (TF) amputation for chronic limb threatening ischemia (CLTI). The overall goal is to enhance prediction of future mobility to assist the rehabilitation team in setting appropriate, evidence-based goals, to better inform patient expectations, and to facilitate appropriate prosthetic prescription. Mixed-methods study that identified patients retrospectively through a large Veteran's Affairs (VA) dataset then prospectively collected patient reported mobility. The Veterans Administration Corporate Data Warehouse (CDW), the National Prosthetics Patient Database (NPPD), participant mailings and phone calls. 357 Veterans who underwent an incident TT or TF amputation for CLTI and received a qualifying lower limb prosthesis (LLP) between March 1, 2018, and November 30, 2020. Not applicable. The Amputee Single Item Mobility Measure (AMPSIMM) was divided into a 4-category outcome to predict wheelchair mobility (0-2), and household (3), basic community (4), or advanced community ambulation (5-6). Variable selection, using a machine learning methodology, led to a final model of 23 predictors that effectively discriminates household ambulation from basic community ambulation and from advanced community ambulation – levels of key clinical importance when estimating future prosthetic demands. The overall model performance was modest as it did not discriminate wheelchair from household mobility as effectively. The AMPREDICT PROsthetics model can assist providers in estimating individual patient's future mobility at the time of prosthetic prescription thereby aiding in the formulation of appropriate mobility goals, as well as facilitating the prescription of a prosthetic device that is most appropriate for their anticipated functional goals.

Prediction of Hepatic Toxicity after Radiotherapy Using a Neural Network Including Blood Biomarkers and Liver Dose Distributions

<h3>Purpose/Objective(s)</h3> ASTRO recently published practical guidelines for treating unresectable hepatocellular carcinoma with radiotherapy, recommending hypofractionated dose escalated regimens when tolerated by the liver. This highlights the need for patient-specific prediction models of radiation-induced hepatic toxicities, which necessitates inclusion not only of radiation dose, but also baseline liver function biomarkers. <h3>Materials/Methods</h3> We developed a shallow neural network to predict declines in Child-Pugh score (CP) of 2 or more, with a focus on robustness, interpretability and ability to include both proton and photon radiation dose distributions. The network uses 1 convolution layer to extract 8 dosimetric features that are representative of the differential DVH (dDVH) and interactions of different dose levels within, which are then combined in a fully connected layer with baseline platelet counts and liver function biomarkers—CP and albumin-bilirubin grade. After registering a study analysis plan, we trained the model on 117 patients from a single institution and independently validated it on 88 patients from another institution, where 47%/36% of the training/validation cohorts received proton radiotherapy. We compared the model to logistic regression to evaluate the importance of feature interactions on predictive power, measured by the area under the receiver operating characteristics (AUC). <h3>Results</h3> The model prediction of CP decline had a cross-validation AUC = 0.86; 95%CI: 0.69-0.97 and independent validation AUC = 0.74 with 89% accuracy in 10% high-risk group. In contrast, logistic regression showed cross-validation AUC 0.73; 95%CI: 0.56-0.88 and validation AUC 0.56, demonstrating a significant importance of feature interactions. Sensitivity analysis of the neural network revealed that patients with low baseline platelet counts were most sensitive to high doses to the normal liver compared to other patient groups, indicating a potential benefit of liver-sparing therapies in these patients. <h3>Conclusion</h3> We developed a patient-specific hepatic toxicity model that can take into account both proton and photon dose distributions and the interaction of radiation dose to normal liver with baseline liver function. Robust model performance across institutions and particularly in high-risk patients could allow stratification of patients to standard, hypofractionated regimens, or proton therapy based on a patient's risk score.

Prediction of Venous Thromboembolism after Ankle Fractures using Machine Learning: To Give Prophylaxis or Not to Give, That is the Question

Category: Ankle; Other Introduction/Purpose: The risk of venous thromboembolism (VTE) after foot and ankle surgery is significantly lower than rates after hip/knee arthroplasty, but it isn't zero. Specific subgroups of patients may be at higher risk, forcing patients and clinicians to navigate the risks and benefits of chemoprophylaxis with insufficient data. Efforts have been made to add clarity to such decision making using conventional data-analysis and risk-scoring methods, but none of these methods were patient-specific or built on robust models of a given patient's individual characteristics. In this study we used machine-learning to determine the potential risk factors for VTE after ankle fracture. We aimed to develop a patient-specific predictive model to assist clinicians and patients in deciding upon the use of postoperative chemoprophylaxis after foot and ankle surgery. Methods: In this preliminary machine-learning-based case-control study, 16,421 patients with ankle fractures were recruited retrospectively. We used an automated-string search method to find patients who were clinically suspected to have developed VTE. Among 1176 such patients, 239 had confirmed VTE within 180 days of sustaining an ankle fracture (cases) and 937 did not (controls). Groups were further sub-divided in patients who had been receiving chemoprophylaxis and those who hadn't. Over 110 factors and variables including patient demographics, past-medical and surgical history, fracture characteristics, treatment, medications, and laboratory values were included in our machine-learning dataset. Three analytical algorithms were used in our machine-learning methods including backward-logistic-regression, decision-tree-classifier (depth=5), and neural networks (two dense layers (n=16 and 4), two drop-out layers, and a sigmoid classifier). Conventional statistics were also used to compare the case and control groups (chi-squared, t-test, p&lt;0.05 considered significant), and the odds-ratio (OR) was calculated for significant parameters. Results: Based on overall performance scores including specificity, sensitivity, area under the ROC curve, accuracy, PPV, NPV, F- 1 score, among the 3 machine-learning methods, the Backward-Logistic-Regression model was superior in predicting VTE post ankle fracture and in determining whether administering prophylaxis can be beneficial for the patient or not (Tables 1 and 2). Other than the previously suggested risk factors, our algorithms showed a positive correlation between the incidence of VTE and smoking (OR=2.09, p&lt;0.001), age &lt;55 y/o (p=0.001), open fracture (OR=2.49, p&lt;0.001), male sex (OR=1.98, p&lt;0.001), surgical versus nonoperative treatment (OR=1.88, p=0.001), and multiple fractures at the time of trauma (OR=1.9, p=0.001). Factors that showed negative correlation with VTE include statins use (OR=0.36, p&lt;0.001), hypertension (OR=0.53, p=0.001), vitamin D (OR=0.43, 0.002), calcium supplementation (OR= 0.43, p= 0.01), hyperlipidemia (OR=0.55, p=0.006), cataract (OR=0.19, p=0.01), osteoporosis (OR=0.36, p=0.02), cardiovascular diseases (OR=0.54, p=0.02), hypokalemia (OR=0.26, p=0.03), and proton pump inhibitor use (OR=0.5, p=0.03). Conclusion: Our machine learning algorithms showed that factors such as tobacco use, younger age, open fracture, multi-trauma, operative treatment, as well as male sex heightened the risk of VTE. In contrast, certain factors such as vitamin D supplementation had negative correlation with VTE. Machine learning algorithms acted in a more complex manner and incorporated more factors in decision-making compared to conventional methods. External validation using larger and more granular datasets as well as using the algorithms in trial modes (shadow modes) are needed to build trust in this algorithm to assist clinicians in predicting/preventing VTE after foot and ankle surgeries.

Comparative Analysis of the Ability of Machine Learning Models in Predicting In-hospital Postoperative Outcomes After Total Hip Arthroplasty.

Machine learning (ML) methods have shown promise in a wide range of applications including the development of patient-specific predictive models before surgical interventions. The purpose of this study was to develop, test, and compare four distinct ML models to predict postoperative parameters after primary total hip arthroplasty. Data from the Nationwide Inpatient Sample were used to identify patients undergoing total hip arthroplasty from 2016 to 2017. Linear support vector machine (LSVM), random forest (RF), neural network (NN), and extreme gradient boost trees (XGBoost) predictive of mortality, length of stay, and discharge disposition were developed and validated using 15 predictive patient-specific and hospital-specific factors. Area under the curve of the receiver operating characteristic (AUCROC) curve and accuracy were used as validity metrics, and the strongest predictive variables under each model were assessed. A total of 177,442 patients were included in this analysis. For mortality, the XGBoost, NN, and LSVM models all had excellent responsiveness during validation while RF had fair responsiveness. LSVM had the highest responsiveness with an AUCROC of 0.973 during validation. For the length of stay, the LSVM and NN models had fair responsiveness while the XGBoost and random forest models had poor responsiveness. LSVM had the highest responsiveness with an AUCROC of 0.744 during validation. For the discharge disposition outcome, LSVM had good responsiveness while the XGBoost, NN, and RF models all had fair responsiveness. LSVM had the highest responsiveness with an AUCROC of 0.801. The ML methods tested demonstrated a range of poor-to-excellent responsiveness and accuracy in the prediction of the assessed metrics, with LSVM being the best performer. Such models should be further developed, with eventual integration into clinical practice to inform patient discussions and management decision making, with the potential for integration into tiered bundled payment models.

Characterization of Puncture Forces of the Human Trachea and Cricothyroid Membrane.

Accurate human tissue biomechanical data represents a critical knowledge gap that will help facilitate the advancement of new medical devices, patient-specific predictive models, and training simulators. Tissues related to the human airway are a top priority, as airway medical procedures are common and critical. Placement of a surgical airway, though less common, is often done in an emergent (cricothyrotomy) or urgent (tracheotomy) fashion. This study is the first to report relevant puncture force data for the human cricothyroid membrane and tracheal annular ligaments. Puncture forces of the cricothyroid membrane and tracheal annular ligaments were collected from 39 and 42 excised human donor tracheas, respectively, with a mechanized load frame holding various surgical tools. The average puncture force of the cricothyroid membrane using an 11 blade scalpel was 1.01 ± 0.36 N, and the average puncture force of the tracheal annular ligaments using a 16 gauge needle was 0.98 ± 0.34 N. This data can be used to inform medical device and airway training simulator development as puncture data of these anatomies has not been previously reported.

Using Artificial Intelligence to Revolutionise the Patient Care Pathway in Hip and Knee Arthroplasty (ARCHERY): Protocol for the Development of a Clinical Prediction Model.

BackgroundHip and knee osteoarthritis is substantially prevalent worldwide, with large numbers of older adults undergoing joint replacement (arthroplasty) every year. A backlog of elective surgery due to the COVID-19 pandemic, and an aging population, has led to substantial issues with access to timely arthroplasty surgery. A potential method to improve the efficiency of arthroplasty services is by increasing the percentage of patients who are listed for surgery from primary care referrals. The use of artificial intelligence (AI) techniques, specifically machine learning, provides a potential unexplored solution to correctly and rapidly select suitable patients for arthroplasty surgery.ObjectiveThis study has 2 objectives: (1) develop a cohort of patients with referrals by general practitioners regarding assessment of suitability for hip or knee replacement from National Health Service (NHS) Grampian data via the Grampian Data Safe Haven and (2) determine the demographic, clinical, and imaging characteristics that influence the selection of patients to undergo hip or knee arthroplasty, and develop a tested and validated patient-specific predictive model to guide arthroplasty referral pathways.MethodsThe AI to Revolutionise the Patient Care Pathway in Hip and Knee Arthroplasty (ARCHERY) project will be delivered through 2 linked work packages conducted within the Grampian Data Safe Haven and Safe Haven Artificial Intelligence Platform. The data set will include a cohort of individuals aged ≥16 years with referrals for the consideration of elective primary hip or knee replacement from January 2015 to January 2022. Linked pseudo-anonymized NHS Grampian health care data will be acquired including patient demographics, medication records, laboratory data, theatre records, text from clinical letters, and radiological images and reports. Following the creation of the data set, machine learning techniques will be used to develop pattern classification and probabilistic prediction models based on radiological images. Supplemental demographic and clinical data will be used to improve the predictive capabilities of the models. The sample size is predicted to be approximately 2000 patients—a sufficient size for satisfactory assessment of the primary outcome. Cross-validation will be used for development, testing, and internal validation. Evaluation will be performed through standard techniques, such as the C statistic (area under curve) metric, calibration characteristics (Brier score), and a confusion matrix.ResultsThe study was funded by the Chief Scientist Office Scotland as part of a Clinical Research Fellowship that runs from August 2021 to August 2024. Approval from the North Node Privacy Advisory Committee was confirmed on October 13, 2021. Data collection started in May 2022, with the results expected to be published in the first quarter of 2024. ISRCTN registration has been completed.ConclusionsThis project provides a first step toward delivering an automated solution for arthroplasty selection using routinely collected health care data. Following appropriate external validation and clinical testing, this project could substantially improve the proportion of referred patients that are selected to undergo surgery, with a subsequent reduction in waiting time for arthroplasty appointments.Trial RegistrationISRCTN Registry ISRCTN18398037; https://www.isrctn.com/ISRCTN18398037International Registered Report Identifier (IRRID)PRR1-10.2196/37092