Intelligence-based medicine最新文献

Purpose

This study aimed to stratify eyes with keratoconus (KC) based on longitudinal changes in all Pentacam parameters into clusters using unsupervised machine learning, with the broader objective of more clearly defining the characteristics of KC progression.

Methods

A data-driven cluster analysis (hierarchical clustering) was undertaken on a retrospective cohort of 1017 kC eyes and 128 control eyes. Clusters were derived using 6-month tomographical change in individual eyes from analysis of the reduced dimensionality parameter space using all available Pentacam parameters (406 principal components). The optimal number of clusters was determined by the clustering's capacity to discriminate progression between KC and control eyes based on change across parameters. One-way ANOVA was used to compare parameters between inferred clusters. Complete Pentacam data changes at 6, 12 and 18-month time points provided validation datasets to determine the generalizability of the clustering model.

Results

We identified three clusters in KC progression patterns. Eyes designated within cluster 3 had the most rapidly changing tomographical parameters compared to eyes in either cluster 1 or 2. Eyes designated within cluster 1 reflected minimal changes in tomographical parameters, closest to the tomographical changes of control (non-KC) eyes. Thirty-nine corneal curvature parameters were identified and associated with these stratified clusters, with each of these parameters changing significantly different between three clusters. Similar clusters were identified at the 6, 12 and 18-month follow-up.

Conclusions

The clustering model developed was able to automatically detect and categorize KC tomographical features into fast, slow, or limited change at different time points. This new KC stratification tool may provide an opportunity to provide a precision medicine approach to KC.

Fall risk assessment can be informed by understanding mobility/gait. Contemporary mobility analysis is being progressed by wearable inertial measurement units (IMU). Typically, IMUs gather temporal mobility-based outcomes (e.g., step time) from labs/clinics or beyond, capturing data for habitually informed fall risk. However, a thorough understanding of free-living IMU-based mobility is currently limited due to a lack of context. For example, although IMU-based length variability can be measured, no absolute clarity exists for factors relating to those variations, which could be due to an intrinsic or an extrinsic environmental factor. For a thorough understanding of habitual-based fall risk assessment through IMU-based mobility outcomes, use of wearable video cameras is suggested. However, investigating video data is laborious i.e., watching and manually labelling environments. Additionally, it raises ethical issues such as privacy. Accordingly, automated artificial intelligence (AI) approaches, that draw upon heterogenous datasets to accurately classify environments, are needed. Here, a novel dataset was created through mining online video and a deep learning-based tool was created via chained convolutional neural networks enabling automated environment (indoor or outdoor) and terrain (e.g., carpet, grass) classification. The dataset contained 146,624 video-based images (environment: 79,251, floor visible: 28,347, terrain: 39,026). Upon training each classifier, the system achieved F1-scores of ≥0.84 when tested on a manually labelled unseen validation dataset (environment: 0.98, floor visible indoor: 0.86, floor visible outdoor: 0.96, terrain indoor: 0.84, terrain outdoor: 0.95). Testing on new data resulted in accuracies from 51 to 100% for isolated networks and 45–90% for complete model. This work is ongoing with the underlying AI being refined for improved classification accuracies to aid automated contextual analysis of mobility/gait and subsequent fall risk. Ongoing work involves primary data capture from within participants free-living environments to bolster dataset heterogeneity.

A growing number of artificial intelligence-based medical devices are receiving clearance from the Food and Drug Administration (FDA). Debate has arisen about best practices for the regulation and safe oversight of such devices whose capabilities, if “unlocked”, include iterative learning and adaptation with exposure to new data. One regulatory mechanism proposed by the FDA is the predetermined change control plan (PCCP). This analysis provides what we believe would be the first example of how a PCCP has been leveraged to improve the performance of a de novo autism diagnostic device in practice. Following the PCCP's model update procedures included in the marketing authorization of the first generation of the device (“algorithm V1”), we conducted an algorithmic threshold optimization procedure to improve the device's ability to detect or rule out autism in children ages 18–72 months without changing the accuracy or intended use of the device. Decision threshold optimization was achieved using a repeated train/test validation procedure on a dataset of 722 children with concern for developmental delay, aged 18–72 months (28% autism, 22% neurotypical, 50% other developmental delay, mean age 3.6 years, 39% female). In 1000 repeats, 70% of samples were selected for threshold optimization and 30% for evaluation, ensuring that no training data appeared in the test set. Out-of-sample performance was estimated by evaluating the selected threshold pair on the test set and comparing the performance metrics of the new pair to the corresponding V1 metrics on the same test set. The device, with optimized decision thresholds, produced a determinate output for 66.5% (95% CI, 62.5–71.0) of children. Positive Predictive Value (PPV) and Negative Predictive Value (PPV) were 87.5% (95% CI, 82.5–96.7) and 95.6% (95% CI, 93.7–97.9) respectively. Threshold optimization improved the device's ability to accurately detect or rule out autism in a greater proportion of children. Given the current waitlist crisis for autism evaluations in the United States, the potential increase in coverage offered by the optimized thresholds is promising and emphasizes the value of regulatory mechanisms that allow software as medical devices to adapt safely and appropriately given real world data.

This paper presents a novel approach for predicting the efficacy of Doxorubicin treatment for glioblastoma. Glioblastomas' rapid growth places them among the most aggressive cancers, killing thousands of Americans every year. The rapid progression of glioblastoma coupled with the high cost of cranial imaging makes clinical decision-making uniquely challenging. Doxorubicin is a commonly used chemotherapy drug to treat glioblastomas. However, predicting the treatment's efficacy remains challenging and time-consuming. Inaccurate predictions can lead to ineffective treatments, severe side effects, and even death. To address this issue, a framework was developed that amalgamates deep learning and differential equations to accurately predict tumor volume growth over time. Specifically, a 2D U-net convolutional neural network (CNN) was employed to segment MRI brain tumor regions and obtain initial volumes. The Gompertz differential equation was then utilized to model the predicted tumor volume growth over time, achieving a mean absolute percent error of 4.98 %. The Gompertz model was modified to incorporate the cytotoxic effect of Doxorubicin treatment. The methodology predicted the final tumor volume of the tumor after being treated with Doxorubicin over multiple 21-day cycles, enabling us to predict the efficacy of treatment and identify patients who may benefit most from this therapy. A user-friendly web application was developed to allow users to input NIFTI files of MRI scans and receive as output a time-course prediction of tumor volume with and without chemotherapy treatment. This approach provides a prediction of Doxorubicin treatment efficacy and can improve patient outcomes and treatment plans.

This work aimed to identify pre-existing health conditions of patients with traumatic brain injury (TBI) and develop predictive models for the first TBI event and its external causes by employing a combination of unsupervised and supervised learning algorithms. We acquired up to five years of pre-injury diagnoses for 488,107 patients with TBI and 488,107 matched control patients who entered the emergency department or acute care hospitals between April 1st, 2002, and March 31st, 2020. Diagnoses were obtained from the Ontario Health Insurance Plan (OHIP) database which contains province-wide claims data by physicians in Ontario, Canada for inpatient and outpatient services. A screening process was conducted on the OHIP diagnostic codes to limit the subsequent analysis to codes that were predictive of TBI, which concluded that 314 codes were significantly associated with TBI. The Latent Dirichlet Allocation (LDA) model was applied to the diagnostic codes and generated an optimal number of 19 topics that concur with published literature but also suggest other unexplored areas. Estimated word-topic probabilities from the LDA model helped us detect pre-morbid conditions among patients with TBI by uncovering the underlying patterns of diagnoses, meanwhile estimated document-topic probabilities were utilized in variable creation as form of a dimension reduction. We created 19 topic scores for each patient in the cohort which were utilized along with socio-demographic factors for Random Forest binary classifier models. Test set performances evaluated using area under the receiver operating characteristic curve (AUC) were: TBI event (AUC = 0.85), external cause of injury: falls (AUC = 0.85), struck by/against (AUC = 0.83), cyclist collision (AUC = 0.76), motor vehicle collision (AUC = 0.83). Our analysis successfully demonstrated the feasibility of using machine learning to predict TBI due to various external causes and identified the most important factors that contribute to this prediction.