Predicting phenotypes of elderly from resting state fMRI

Machine learning techniques are increasingly embraced in neuroimaging studies of healthy and diseased human brains. They have been used successfully in predicting phenotypes, or even clinical outcomes, and in turning functional connectome metrics into phenotype biomarkers of both healthy individuals and patients. In this study, we used functional connectivity characteristics based on resting state functional magnetic resonance imaging data to accurately classify healthy elderly in terms of their phenotype status. Additionally, as the functional connections that contribute to the classification can be identified, we can draw inferences about the network that is predictive of the investigated phenotypes. Our proposed pipeline for phenotype classification can be expanded to other phenotypes (cognitive, psychological, clinical) and possibly be used to shed light on the modifiable risk and protective factors in normative and pathological brain aging.

Microcystin concentrations and congener composition in relation to environmental variables across 440 north-temperate and boreal lakes

Understanding the environmental conditions and taxa that promote the occurrence of cyanobacterial toxins is imperative for effective management of lake ecosystems. Herein, we modeled total microcystin presence and concentrations with a broad suite of environmental predictors and cyanobacteria community data collected across 440 Canadian lakes using standardized methods. We also conducted a focused analysis targeting 14 microcystin congeners across 190 lakes, to examine how abiotic and biotic factors influence their relative proportions. Microcystins were detected in 30 % of lakes, with the highest total concentrations occurring in the most eutrophic lakes located in ecozones of central Canada. The two most commonly detected congeners were MC-LR (61 % of lakes) and MC-LA (37 % of lakes), while 11 others were detected more sporadically across waterbodies. Congener diversity peaked in central Canada where cyanobacteria biomass was highest. Using a zero-altered hurdle model, the probability of detecting microcystin was best explained by increasing Microcystis biomass, Daphnia and cyclopoid biomass, soluble reactive phosphorus, pH and wind. Microcystin concentrations increased with the biomass of Microcystis and other less dominant cyanobacteria taxa, as well as total phosphorus, cyclopoid copepod biomass, dissolved inorganic carbon and water temperature. Collectively, these models accounted for 34 % and 70 % of the variability, respectively. Based on a multiple factor analysis of microcystin congeners, cyanobacteria community data, environmental and zooplankton data, we found that the relative abundance of most congeners varied according to trophic state and were related to a combination of cyanobacteria genera biomasses and environmental variables.

Computerized Differentiation of Growth Status for Abdominal Aortic Aneurysms: A Feasibility Study

Fast-growing abdominal aortic aneurysms (AAA) have a high rupture risk and poor outcomes if not promptly identified and treated. Our primary objective is to improve the differentiation of small AAAs' growth status (fast versus slow-growing) through a combination of patient health information, computational hemodynamics, geometric analysis, and artificial intelligence. 3D computed tomography angiography (CTA) data available for 70 patients diagnosed with AAAs with known growth status were used to conduct geometric and hemodynamic analyses. Differences among ten metrics (out of ninety metrics) were statistically significant discriminators between fast and slow-growing groups. Using a support vector machine (SVM) classifier, the area under receiving operating curve (AUROC) and total accuracy of our best predictive model for differentiation of AAAs' growth status were 0.86 and 77.50%, respectively. In summary, the proposed analytics has the potential to differentiate fast from slow-growing AAAs, helping guide resource allocation for the management of patients with AAAs.

Spatial and temporal variability of mercury in Upper and Lower Red Lake Walleye

Mercury is a global pollutant that is released into our environment by natural and anthropogenic processes resulting in extensive studies of mercury cycling in aquatic ecosystems, and the issuance of human-health-based fish-consumption advisories. We examined total mercury concentrations in Walleye Sander vitreus from Upper and Lower Red Lakes, located in north central Minnesota, between 2019 and 2020. Sampled Walleye (n = 265) ranged from 158 to 610 mm in total length from an age range of young-of-the year to 16 years. Mercury concentrations within the Walleye ranged from 0.030 mg/kg to 0.564 mg/kg (x̄ = 0.179 ± 0.105 mg/kg; x̄ = mean ± sd, all fish-mercury concentrations expressed on wet-weight basis). The best supported model for predicting mercury concentrations in Red Lake Walleye included the independent variables: length, age, sex, and lake basin. This model indicated that there was a significant difference in mercury concentrations between Upper and Lower Red Lake (x̄ = 0.215 ± 0.117 and 0.144 ± 0.077 mg/kg, respectively), and also suggests that individuals who rely on fish for subsistence should target Walleye that are ≤ 400 mm from Lower Red Lake. Observed differences in mercury concentrations could be linked to wetland area influences, fish growth rates, and physicochemical parameters between the two basins. Given that our results illustrated a significant difference in fish-mercury concentrations between basins, future pollutant monitoring efforts should treat Upper and Lower Red Lake as separate lakes and not assume that data from one basin can apply to the other.

Computational neurosciences and quantitative systems pharmacology: a powerful combination for supporting drug development in neurodegenerative diseases

Successful clinical development of new therapeutic interventions is notoriously difficult, especially in neurodegenerative diseases, where predictive biomarkers are scarce and functional improvement is often based on patient's perception, captured by structured interviews. As a consequence, mechanistic modeling of the processes relevant to therapeutic interventions in CNS disorders has been lagging behind other disease indications, probably because of the perceived complexity of the brain. However in this report, we develop the argument that a combination of Computational Neurosciences and Quantitative Systems Pharmacology (QSP) modeling of molecular pathways is a powerful simulation tool to enhance the probability of successful drug development for neurodegenerative diseases. Computational Neurosciences aims to predict action potential dynamics and neuronal circuit activation that are ultimately linked to behavioral changes and clinically relevant functional outcomes. These processes can not only be affected by the disease state, but also by common genotype variants on neurotransmitter-related proteins and the psycho-active medications often prescribed in these patient populations. Quantitative Systems Pharmacology (QSP) modeling of molecular pathways allows to simulate key pathological drivers of dementia, such as protein aggregation and neuroinflammatory responses. They often impact neurotransmitter homeostasis and voltage-gated ion-channels or lead to mitochondrial dysfunction, ultimately leading to changes in action potential dynamics and clinical readouts. Combining these two modeling approaches can lead to better actionable understanding of the many non-linear pharmacodynamic processes active in the human diseased brain. Practical applications include a rational selection of the optimal doses in combination therapies, identification of subjects more likely to respond to treatment, a more balanced stratification of treatment arms in terms of comedications, disease status and common genotype variants and re-analysis of small clinical trials to uncover a possible clinical signal. Ultimately this will lead to a higher success rate of bringing new therapeutics to the right patient populations.

Objective wearable measures and subjective questionnaires for predicting response to neurostimulation in people with chronic pain

BACKGROUND

Neurostimulation is an effective therapy for treating and management of refractory chronic pain. However, the complex nature of pain and infrequent in-clinic visits, determining subject's long-term response to the therapy remains difficult. Frequent measurement of pain in this population can help with early diagnosis, disease progression monitoring, and evaluating long-term therapeutic efficacy. This paper compares the utilization of the common subjective patient-reported outcomes with objective measures captured through a wearable device for predicting the response to neurostimulation therapy.

METHOD

Data is from the ongoing international prospective post-market REALITY clinical study, which collects long-term patient-reported outcomes from 557 subjects implanted by Spinal Cord Stimulator (SCS) or Dorsal Root Ganglia (DRG) neurostimulators. The REALITY sub-study was designed for collecting additional wearables data on a subset of 20 participants implanted with SCS devices for up to six months post implantation. We first implemented a combination of dimensionality reduction algorithms and correlation analyses to explore the mathematical relationships between objective wearable data and subjective patient-reported outcomes. We then developed machine learning models to predict therapy outcome based on the subject's response to the numerical rating scale (NRS) or patient global impression of change (PGIC).

RESULTS

Principal component analysis showed that psychological aspects of pain were associated with heart rate variability, while movement-related measures were strongly associated with patient-reported outcomes related to physical function and social role participation. Our machine learning models using objective wearable data predicted PGIC and NRS outcomes with high accuracy without subjective data. The prediction accuracy was higher for PGIC compared with the NRS using subjective-only measures primarily driven by the patient satisfaction feature. Similarly, the PGIC questions reflect an overall change since the study onset and could be a better predictor of long-term neurostimulation therapy outcome.

CONCLUSIONS

The significance of this study is to introduce a novel use of wearable data collected from a subset of patients to capture multi-dimensional aspects of pain and compare the prediction power with the subjective data from a larger data set. The discovery of pain digital biomarkers could result in a better understanding of the patient's response to therapy and their general well-being.

Predicting post-discharge venous thromboembolism and bleeding among medical patients: External validation of a novel risk score utilizing ubiquitous biomarkers

BACKGROUND

Post-hospitalization thromboprophylaxis can reduce venous thromboembolism (VTE) risk for non-surgical patients but may carry bleeding risks. We aimed to externally validate the Intermountain Risk Scores for hospital-associated venous thromboembolism (HA-VTE IMRS) and major bleeding (HA-MB IMRS) for VTE and bleeding outcomes.

METHODS

Retrospective cohort study of adult patients discharged alive from medical services between 2015 and 2019. HA-VTE IMRS and HA-MB IMRS were calculated at the time of hospital discharge and dichotomized as high- or low-risk as described in the derivation manuscript. 90-day post-discharge VTE outcomes were assessed from diagnostic radiology reports, and bleeding outcomes were assessed using ICD-10 codes and blood bank transfusion records.

RESULTS

Among 113,578 patients in the study, 66,340 patients (58.4 %) had a low-risk HA-VTE IMRS <7, versus 47,238 (41.6 %) high-risk ≥7. For bleed prediction, 71,576 patients (63 %) had a low-risk HA-MB IMRS <8, versus 42,002 (37 %) high-risk ≥8. VTE incidence was 1.1 % and 0.6 % while major bleeding incidence was 1.3 % and 0.1 % in high-risk versus low-risk cohorts, respectively. AUCs for VTE and bleed outcome discrimination were 0.59 and 0.78, respectively. Patients with a combined high-risk VTE score and low-risk bleeding score comprised 14.5 % of the population.

CONCLUSION

In this external validation study, the HA-VTE IMRS had poor discrimination for VTE but the HA-MB IMRS had good discriminatory ability for major bleeding events. A sizable minority of patients were categorized as high VTE risk with low bleed risk, a population which may have an optimal risk-benefit profile for post-hospital thromboprophylaxis.

Growth behavior of Shiga toxin-producing Escherichia coli, Salmonella, and generic E. coli in raw pork considering background microbiota at 10, 25, and 40 °C

Recent epidemiological evidence suggests that pork products may be vehicles for the transmission of Shiga toxin-producing Escherichia coli (STEC) to humans. The severe morbidity associated with STEC infections highlights the need for research to understand the growth behavior of these bacteria in pork products. Classical predictive models can estimate pathogen growth in sterile meat. However, competition models considering background microbiota reflect a more realistic scenario for raw meat products. The objective of this study was to estimate the growth kinetics of clinically significant STEC (O157, non-O157, and O91), Salmonella, and generic E. coli in raw ground pork using competition primary growth models at temperature abuse (10 and 25 °C) and sublethal temperature (40 °C). A competition model incorporating the No lag Buchanan model was validated using the acceptable prediction zone (APZ) method where >92 % (1498/1620) of the residual errors fell within the APZ (pAPZ > 0.70). The background microbiota (mesophilic aerobic plate counts, APC) inhibited the growth of STEC and Salmonella indicating a simple one-directional competitive interaction between pathogens and the mesophilic microbiota of ground pork. The maximum specific growth rate (μ_max) of all the bacterial groups was not significantly different (p > 0.05) based on fat content (5 vs 25 %) except for generic E. coli at 10 °C. E. coli O157 and non-O157 behaved similarly in terms of μ_max and maximum population density (MPD). Salmonella showed a similar (p > 0.05) μ_max to E. coli O157 and non-O157 at 10 and 40 °C but a significantly higher rate (p < 0.05) at 25 °C. STEC were more prone to be inhibited by APC than Salmonella at 10 and 25 °C. The μ_max of O91 was lower (p < 0.05) than other STEC and Salmonella at 10 and 25 °C but similar (p > 0.05) at 40 °C. Generic E. coli showed a two- to five-times higher (p < 0.05) μ_max (0.028 ± 0.011 log₁₀ CFU/h) than other bacterial groups (0.006 ± 0.004 to 0.012 ± 0.003 log₁₀ CFU/h) at 10 °C making it a potential indicator bacteria for process control. Industry and regulators can use competitive models to develop appropriate risk assessment and mitigation strategies to improve the microbiological safety of raw pork products.

One Size Does Not Fit All: Methodological Considerations for Brain-Based Predictive Modeling in Psychiatry

Psychiatric illnesses are heterogeneous in nature. No illness manifests in the same way across individuals, and no two patients with a shared diagnosis exhibit identical symptom profiles. Over the last several decades, group-level analyses of in vivo neuroimaging data have led to fundamental advances in our understanding of the neurobiology of psychiatric illnesses. More recently, access to computational resources and large, publicly available datasets alongside the rise of predictive modeling and precision medicine approaches have facilitated the study of psychiatric illnesses at an individual level. Data-driven machine learning analyses can be applied to identify disease-relevant biological subtypes, predict individual symptom profiles, and recommend personalized therapeutic interventions. However, when developing these predictive models, methodological choices must be carefully considered to ensure accurate, robust, and interpretable results. Choices pertaining to algorithms, neuroimaging modalities and states, data transformation, phenotypes, parcellations, sample sizes, and populations we are specifically studying can influence model performance. Here, we review applications of neuroimaging-based machine learning models to study psychiatric illnesses and discuss the effects of different methodological choices on model performance. An understanding of these effects is crucial for the proper implementation of predictive models in psychiatry and will facilitate more accurate diagnoses, prognoses, and therapeutics.

Task fMRI paradigms may capture more behaviorally relevant information than resting-state functional connectivity

Characterizing the optimal fMRI paradigms for detecting behaviorally relevant functional connectivity (FC) patterns is a critical step to furthering our knowledge of the neural basis of behavior. Previous studies suggested that FC patterns derived from task fMRI paradigms, which we refer to as task-based FC, are better correlated with individual differences in behavior than resting-state FC, but the consistency and generalizability of this advantage across task conditions was not fully explored. Using data from resting-state fMRI and three fMRI tasks from the Adolescent Brain Cognitive Development Study ® (ABCD), we tested whether the observed improvement in behavioral prediction power of task-based FC can be attributed to changes in brain activity induced by the task design. We decomposed the task fMRI time course of each task into the task model fit (the fitted time course of the task condition regressors from the single-subject general linear model) and the task model residuals, calculated their respective FC, and compared the behavioral prediction performance of these FC estimates to resting-state FC and the original task-based FC. The FC of the task model fit was better than the FC of the task model residual and resting-state FC at predicting a measure of general cognitive ability or two measures of performance on the fMRI tasks. The superior behavioral prediction performance of the FC of the task model fit was content-specific insofar as it was only observed for fMRI tasks that probed similar cognitive constructs to the predicted behavior of interest. To our surprise, the task model parameters, the beta estimates of the task condition regressors, were equally if not more predictive of behavioral differences than all FC measures. These results showed that the observed improvement of behavioral prediction afforded by task-based FC was largely driven by the FC patterns associated with the task design. Together with previous studies, our findings highlighted the importance of task design in eliciting behaviorally meaningful brain activation and FC patterns.

Multimodal predictions of treatment outcome in major depression: A comparison of data-driven predictors with importance ratings by clinicians

BACKGROUND

Reliable prediction models of treatment outcome in Major Depressive Disorder (MDD) are currently lacking in clinical practice. Data-driven outcome definitions, combining data from multiple modalities and incorporating clinician expertise might improve predictions.

METHODS

We used unsupervised machine learning to identify treatment outcome classes in 1060 MDD inpatients. Subsequently, classification models were created on clinical and biological baseline information to predict treatment outcome classes and compared to the performance of two widely used classical outcome definitions. We also related the findings to results from an online survey that assessed which information clinicians use for outcome prognosis.

RESULTS

Three and four outcome classes were identified by unsupervised learning. However, data-driven outcome classes did not result in more accurate prediction models. The best prediction model was targeting treatment response in its standard definition and reached accuracies of 63.9 % in the test sample, and 59.5 % and 56.9 % in the validation samples. Top predictors included sociodemographic and clinical characteristics, while biological parameters did not improve prediction accuracies. Treatment history, personality factors, prior course of the disorder, and patient attitude towards treatment were ranked as most important indicators by clinicians.

LIMITATIONS

Missing data limited the power to identify biological predictors of treatment outcome from certain modalities.

CONCLUSIONS

So far, the inclusion of available biological measures in addition to psychometric and clinical information did not improve predictive value of the models, which was overall low. Optimized biomarkers, stratified predictions and the inclusion of clinical expertise may improve future prediction models.

Derivation and validation of predictive indices for cardiac readmission after coronary and valvular surgery - A multicenter study

Objective

To derive and validate models to predict the risk of a cardiac readmission within one year after specific cardiac surgeries using information that is commonly available from hospital electronic medical records.

Methods

In this retrospective cohort study, we derived and externally validated clinical models to predict the likelihood of cardiac readmissions within one-year of isolated CABG, AVR, and combined CABG+AVR in Ontario, Canada, using multiple clinical registries and routinely collected administrative databases. For all adult patients who underwent these procedures, multiple Fine and Gray subdistribution hazard models were derived within a competing-risk framework using the cohort from April 2015 to March 2018 and validated in an independent cohort (April 2018 to March 2020).

Results

For the model that predicted post-CABG cardiac readmission, the c-statistic was 0.73 in the derivation cohort and 0.70 in the validation cohort at one-year. For the model that predicted post-AVR cardiac readmission, the c-statistic was 0.74 in the derivation and 0.73 in the validation cohort at one-year. For the model that predicted cardiac readmission following CABG+AVR, the c-statistic was 0.70 in the derivation and 0.66 in the validation cohort at one-year.

Conclusions

Prediction of one-year cardiac readmission for isolated CABG, AVR, and combined CABG+AVR can be achieved parsimoniously using multidimensional data sources. Model discrimination was better than existing models derived from single and multicenter registries.

Development and internal validation of machine learning algorithms for predicting complications after primary total hip arthroplasty

INTRODUCTION

Complications after total hip arthroplasty (THA) may result in readmission or reoperation and impose a significant cost on the healthcare system. Understanding which patients are at-risk for complications can potentially allow for targeted interventions to decrease complication rates through pursuing preoperative health optimization. The purpose of the current was to develop and internally validate machine learning (ML) algorithms capable of performing patient-specific predictions of all-cause complications within two years of primary THA.

METHODS

This was a retrospective case-control study of clinical registry data from 616 primary THA patients from one large academic and two community hospitals. The primary outcome was all-cause complications at a minimum of 2-years after primary THA. Recursive feature elimination was applied to identify preoperative variables with the greatest predictive value. Five ML algorithms were developed on the training set using tenfold cross-validation and internally validated on the independent testing set of patients. Algorithms were assessed by discrimination, calibration, Brier score, and decision curve analysis to quantify performance.

RESULTS

The observed complication rate was 16.6%. The stochastic gradient boosting model achieved the best performance with an AUC = 0.88, calibration intercept = 0.1, calibration slope = 1.22, and Brier score = 0.09. The most important factors for predicting complications were age, drug allergies, prior hip surgery, smoking, and opioid use. Individual patient-level explanations were provided for the algorithm predictions and incorporated into an open access digital application: https://sorg-apps.shinyapps.io/tha_complication/ CONCLUSIONS: The stochastic boosting gradient algorithm demonstrated good discriminatory capacity for identifying patients at high-risk of experiencing a postoperative complication and proof-of-concept for creating office-based applications from ML that can perform real-time prediction. However, this clinical utility of the current algorithm is unknown and definitions of complications broad. Further investigation on larger data sets and rigorous external validation is necessary prior to the assessment of clinical utility with respect to risk-stratification of patients undergoing primary THA.

LEVEL OF EVIDENCE

III, therapeutic study.

Child-therapist interaction features impact Autism treatment response trajectories

BACKGROUND

Identifying mechanisms of change in Autism treatment may help explain response variability and maximize efficacy. For this, the child-therapist interaction could have a key role as stressed by developmental models of intervention, but still remains under-investigated.

AIMS

The longitudinal study of treatment response trajectories considering both baseline and child-therapist interaction features by means of predictive modeling.

METHODS AND PROCEDURES

N = 25 preschool children were monitored for one year during Naturalistic Developmental Behavioral Intervention. N = 100 video-recorded sessions were annotated with an observational coding system at four time points, to extract quantitative interaction features.

OUTCOMES AND RESULTS

Baseline and interaction variables were combined to predict response trajectories at one year, and achieved the best predictive performance. The baseline developmental gap, therapist's efficacy in child engagement, respecting children's timing after fast behavioral synchronization, and modulating the interplay to prevent child withdrawal emerged as key factors. Further, changes in interaction patterns in the early phase of the intervention were predictive of the overall response to treatment.

CONCLUSIONS AND IMPLICATIONS

Clinical implications are discussed, stressing the importance of promoting emotional self-regulation during intervention and the possible relevance of the first period of intervention for later response.

Improving Resource Utilization for Arthroplasty Care by Leveraging Machine Learning and Optimization: A Systematic Review

Background

There is a growing demand for total joint arthroplasty (TJA) surgery. The applications of machine learning (ML), mathematical optimization, and computer simulation have the potential to improve efficiency of TJA care delivery through outcome prediction and surgical scheduling optimization, easing the burden on health-care systems. The purpose of this study was to evaluate strategies using advances in analytics and computational modeling that may improve planning and the overall efficiency of TJA care.

Methods

A systematic review including MEDLINE, Embase, and IEEE Xplore databases was completed from inception to October 3, 2022, for identification of studies generating ML models for TJA length of stay, duration of surgery, and hospital readmission prediction. A scoping review of optimization strategies in elective surgical scheduling was also conducted.

Results

Twenty studies were included for evaluating ML predictions and 17 in the scoping review of scheduling optimization. Among studies generating linear or logistic control models alongside ML models, only 1 found a control model to outperform its ML counterpart. Furthermore, neural networks performed superior to or at the same level as conventional ML models in all but 1 study. Implementation of mathematical and simulation strategies improved the optimization efficiency when compared to traditional scheduling methods at the operational level.

Conclusions

High-performing predictive ML-based models have been developed for TJA, as have mathematical strategies for elective surgical scheduling optimization. By leveraging artificial intelligence for outcome prediction and surgical optimization, there exist greater opportunities for improved resource utilization and cost-savings in TJA than when using traditional modeling and scheduling methods.

Predicting Workplace Violence in the Emergency Department Based on Electronic Health Record Data

INTRODUCTION

Emergency departments are extremely vulnerable to workplace violence, and emergency nurses are frequently exposed to workplace violence. We developed workplace violence prediction models using machine learning methods based on data from electronic health records.

METHODS

This study was conducted using electronic health record data collected between January 1, 2016 and December 31, 2021. Workplace violence cases were identified based on violence-related mentions in nursing records. Workplace violence was predicted using various factors related to emergency department visit and stay.

RESULTS

The dataset included 1215 workplace violence cases and 6044 nonviolence cases. Random Forest showed the best performance among the algorithms adopted in this study. Workplace violence was predicted with higher accuracy when both ED visit and ED stay factors were used as predictors (0.90, 95% confidence interval 0.898-0.912) than when only ED visit factors were used. When both ED visit and ED stay factors were included for prediction, the strongest predictor of risk of WPV was patient dissatisfaction, followed by high average daily length of stay, high daily number of patients, and symptoms of psychiatric disorders.

DISCUSSION

This study showed that workplace violence could be predicted with previous data regarding ED visits and stays documented in electronic health records. Timely prediction and mitigation of workplace violence could improve the safety of emergency nurses and the quality of nursing care. To prevent workplace violence, emergency nurses must recognize and continuously observe the risk factors for workplace violence from admission to discharge.

Artificial intelligence to predict outcomes of head and neck radiotherapy

Head and neck radiotherapy induces important toxicity, and its efficacy and tolerance vary widely across patients. Advancements in radiotherapy delivery techniques, along with the increased quality and frequency of image guidance, offer a unique opportunity to individualize radiotherapy based on imaging biomarkers, with the aim of improving radiation efficacy while reducing its toxicity. Various artificial intelligence models integrating clinical data and radiomics have shown encouraging results for toxicity and cancer control outcomes prediction in head and neck cancer radiotherapy. Clinical implementation of these models could lead to individualized risk-based therapeutic decision making, but the reliability of the current studies is limited. Understanding, validating and expanding these models to larger multi-institutional data sets and testing them in the context of clinical trials is needed to ensure safe clinical implementation. This review summarizes the current state of the art of machine learning models for prediction of head and neck cancer radiotherapy outcomes.

Artificial intelligence and machine learning overview in pathology & laboratory medicine: A general review of data preprocessing and basic supervised concepts

Machine learning (ML) is becoming an integral aspect of several domains in medicine. Yet, most pathologists and laboratory professionals remain unfamiliar with such tools and are unprepared for their inevitable integration. To bridge this knowledge gap, we present an overview of key elements within this emerging data science discipline. First, we will cover general, well-established concepts within ML, such as data type concepts, data preprocessing methods, and ML study design. We will describe common supervised and unsupervised learning algorithms and their associated common machine learning terms (provided within a comprehensive glossary of terms that are discussed within this review). Overall, this review will offer a broad overview of the key concepts and algorithms in machine learning, with a focus on pathology and laboratory medicine. The objective is to provide an updated useful reference for those new to this field or those who require a refresher.

A pilot study to predict cardiac arrest in the pediatric intensive care unit

BACKGROUND

Cardiac arrest is a leading cause of mortality prior to discharge for children admitted to the pediatric intensive care unit. To address this problem, we used machine learning to predict cardiac arrest up to three hours in advance.

METHODS

Our data consists of 240 Hz ECG waveform data, 0.5 Hz physiological time series data, medications, and demographics from 1,145 patients in the pediatric intensive care unit at the Johns Hopkins Hospital, 15 of whom experienced a cardiac arrest. The data were divided into training, validating, and testing sets, and features were generated every five minutes. 23 heart rate variability (HRV) metrics were determined from ECG waveforms. 96 summary statistics were calculated for 12 vital signs, such as respiratory rate and blood pressure. Medications were classified into 42 therapeutic drug classes. Binary features were generated to indicate the administration of these different drugs. Next, six machine learning models were evaluated: logistic regression, support vector machine, random forest, XGBoost, LightGBM, and a soft voting ensemble.

RESULTS

XGBoost performed the best, with 0.971 auROC, 0.797 auPRC, 99.5% sensitivity, and 69.6% specificity on an independent test set.

CONCLUSION

We have created high-performing models that identify signatures of in-hospital cardiac arrest (IHCA) that may not be evident to clinicians. These signatures include a combination of heart rate variability metrics, vital signs data, and therapeutic drug classes. These machine learning models can predict IHCA up to three hours prior to onset with high performance, allowing clinicians to intervene earlier, improving patient outcomes.

Predictive modeling of indoor dust lead concentrations: Sources, risks, and benefits of intervention

Lead (Pb) contamination continues to contribute to world-wide morbidity in all countries, particularly low- and middle-income countries. Despite its continued widespread adverse effects on global populations, particularly children, accurate prediction of elevated household dust Pb and the potential implications of simple, low-cost household interventions at national and global scales have been lacking. A global dataset (∼40 countries, n = 1951) of community sourced household dust samples were used to predict whether indoor dust was elevated in Pb, expanding on recent work in the United States (U.S.). Binned housing age category alone was a significant (p < 0.01) predictor of elevated dust Pb, but only generated effective predictive accuracy for England and Australia (sensitivity of ∼80%), similar to previous results in the U.S. This likely reflects comparable Pb pollution legacies between these three countries, particularly with residential Pb paint. The heterogeneity associated with Pb pollution at a global scale complicates the predictive accuracy of our model, which is lower for countries outside England, the U.S., and Australia. This is likely due to differing environmental Pb regulations, sources, and the paucity of dust samples available outside of these three countries. In England, the U.S., and Australia, simple, low-cost household intervention strategies such as vacuuming and wet mopping could conservatively save 70 billion USD within a four-year period based on our model. Globally, up to 1.68 trillion USD could be saved with improved predictive modeling and primary intervention to reduce harmful exposure to Pb dust sources.

A practical guide to the development and deployment of deep learning models for the Orthopedic surgeon: part I

Deep learning has a profound impact on daily life. As Orthopedics makes use of this rapid escalation in technology, Orthopedic surgeons will need to take leadership roles on deep learning projects. Moreover, surgeons must possess an understanding of what is necessary to design and implement deep learning-based project pipelines. This review provides a practical guide for the Orthopedic surgeon to understand the steps needed to design, develop, and deploy a deep learning pipeline for clinical applications. A detailed description of the processes involved in defining the problem, building the team, acquiring and curating the data, labeling the data, establishing the ground truth, pre-processing and augmenting the data, and selecting the required hardware is provided. In addition, an overview of unique considerations involved in the training and evaluation of deep learning models is provided. This review strives to provide surgeons with the groundwork needed to identify gaps in the clinical landscape that deep learning models may be able to fill and equips them with the knowledge needed to lead an interdisciplinary team through the process of creating novel deep-learning-based solutions to fill those gaps.

Radiographic findings involved in knee osteoarthritis progression are associated with pain symptom frequency and baseline disease severity: a population-level analysis using deep learning

PURPOSE

To (1) develop a deep-learning (DL) algorithm capable of producing limb-length and knee-alignment measurements, and (2) determine the association between limb-length discrepancy (LLD), coronal-plane alignment, osteoarthritis (OA) severity, and patient-reported knee pain.

METHODS

A multicenter, prospective patient cohort from the Osteoarthritis Initiative between 2004 and 2015 with full-limb standing radiographs at 12 month follow-up was included. A convolutional neural network was developed to automate measurements of the hip-knee-ankle (HKA) angle, femur, and tibia lengths, and LLD. At 12 month follow-up, patients reported their frequency of knee pain since enrollment and current level of knee pain.

RESULTS

A total of 1011 patients (2022 knees, 52.3% female) with an average age of 61.2 ± 9.0 years were included. The algorithm performed 12,312 measurements in 5.4 h. ICC values of HKA and LLD ranged between 0.87 and 1.00 when compared against trained radiologist measurements. Knees producing pain most days of the month were significantly more varus (mean HKA:- 3.9° ± 2.8°) or valgus (mean HKA:2.8° ± 2.3°) compared to knees that did not produce any pain (p < 0.05). In varus knees, those producing pain on most days were part of the shorter limb compared to nonpainful knees (p < 0.05). Baseline Kellgren-Lawrence grade was significantly associated with HKA magnitude, LLD, and pain frequency at 12 month follow-up (p < 0.05 all).

CONCLUSION

A higher frequency of knee pain was associated with more severe coronal plane deformity, with valgus deviation being one degree less than varus on average, suggesting that the knee tolerates less valgus deformation before symptoms become more consistent. Knee pain frequency was also associated with greater LLD and baseline KL grade, suggesting an association between radiographically apparent joint degeneration and pain frequency.

LEVEL OF EVIDENCE

IV case series.

Comparison of State-of-the-Art Neural Network Survival Models with the Pooled Cohort Equations for Cardiovascular Disease Risk Prediction

BACKGROUND

The Pooled Cohort Equations (PCEs) are race- and sex-specific Cox proportional hazards (PH)-based models used for 10-year atherosclerotic cardiovascular disease (ASCVD) risk prediction with acceptable discrimination. In recent years, neural network models have gained increasing popularity with their success in image recognition and text classification. Various survival neural network models have been proposed by combining survival analysis and neural network architecture to take advantage of the strengths from both. However, the performance of these survival neural network models compared to each other and to PCEs in ASCVD prediction is unknown.

METHODS

In this study, we used 6 cohorts from the Lifetime Risk Pooling Project (with 5 cohorts as training/internal validation and one cohort as external validation) and compared the performance of the PCEs in 10-year ASCVD risk prediction with an all two-way interactions Cox PH model (Cox PH-TWI) and three state-of-the-art neural network survival models including Nnet-survival, Deepsurv, and Cox-nnet. For all the models, we used the same 7 covariates as used in the PCEs. We fitted each of the aforementioned models in white females, white males, black females, and black males, respectively. We evaluated models' internal and external discrimination power and calibration.

RESULTS

The training/internal validation sample comprised 23216 individuals. The average age at baseline was 57.8 years old (SD = 9.6); 16% developed ASCVD during average follow-up of 10.50 (SD = 3.02) years. Based on 10 × 10 cross-validation, the method that had the highest C-statistics was Deepsurv (0.7371) for white males, Deepsurv and Cox PH-TWI (0.7972) for white females, PCE (0.6981) for black males, and Deepsurv (0.7886) for black females. In the external validation dataset, Deepsurv (0.7032), Cox-nnet (0.7282), PCE (0.6811), and Deepsurv (0.7316) had the highest C-statistics for white male, white female, black male, and black female population, respectively. Calibration plots showed that in 10 × 10 validation, all models had good calibration in all race and sex groups. In external validation, all models overestimated the risk for 10-year ASCVD.

CONCLUSIONS

We demonstrated the use of the state-of-the-art neural network survival models in ASCVD risk prediction. Neural network survival models had similar if not superior discrimination and calibration compared to PCEs.

Predicting Fintech Innovation Adoption: the Mediator Role of Social Norms and Attitudes

Digital innovation is challenging the traditional way of offering financial services to companies; the so-called Fintech phenomenon refers to startups that use the latest technologies to offer innovative financial services. Within the framework of the Theory of Planned Behavior (TPB) and the Theory of Reasoned Action (TRA), the primary purpose of this paper is to develop a causal-predictive analysis of the relationship between Subjective Norms, Attitudes, and Perceived Behavioral Control with the Intention to Use and Behavioral Use of the Fintech services by companies. Partial Least Squares Structural Equation Modeling methodology was used with data collected from a survey of 300 companies. Our findings support the TRA and TPB models and confirm their robustness in predicting companies' intention and use of Fintech services. Financial technology innovators must understand the processes involved in users' adoption to design sound strategies that increase the viability of their services. Studying the antecedents of behavioral intention to adopt Fintech services can greatly help understand the pace of adoption, allowing these players to attract and retain customers better. This study contributes to the literature by formulating and validating TPB to predict Fintech adoption, and its findings provide useful information for banks and Fintech companies and lead to an improvement in organizational performance management in formulating marketing strategies.

Evaluation of a Machine Learning-Based Dysphagia Prediction Tool in Clinical Routine: A Prospective Observational Cohort Study

Based on a large number of pre-existing documented electronic health records (EHR), we developed a machine learning (ML) algorithm for detection of dysphagia and aspiration pneumonia. The aim of our study was to prospectively apply this algorithm in two large patient cohorts. The tool was integrated in the hospital information system of a secondary care hospital in Austria. Based on existing data such as diagnoses, laboratory, and medication, dysphagia risk was predicted automatically, and patients were stratified into three risk groups. Patients' risk groups and risk factors were visualized in a web application. Prospective predictions of 1270 admissions to geriatric or internal medicine departments were compared with the occurrence of dysphagia or aspiration pneumonia of routinely documented events. The discriminative performance for internal medicine patients (n = 885) was excellent with an AUROC of 0.841, a sensitivity of 74.2%, and a specificity of 84.1%. For the smaller geriatric cohort (n = 221), the AUROC was 0.758, sensitivity 44.4%, and specificity 93.0%. For both cohorts, calibration plots showed a slight overestimation of the risk. This is the first study to evaluate the performance of a ML-based prediction tool for dysphagia in a prospective clinical setting. Future studies should validate the predictions on data of systematic dysphagia screening by specialists and evaluate user satisfaction and acceptance. The ML-based dysphagia prediction tool achieved an excellent performance in the internal medicine cohort. More data are needed to determine the performance in geriatric patients.