Saliency methods are designed to provide explainability for deep image processing models by assigning feature-wise importance scores and thus detecting informative regions in the input images. Recently, these methods have been widely adapted to the time series domain, aiming to identify important temporal regions in a time series. This paper extends our former work on identifying the systematic failure of such methods in the time series domain to produce relevant results when informative patterns are based on underlying latent information rather than temporal regions. First, we both visually and quantitatively assess the quality of explanations provided by multiple state-of-the-art saliency methods, including Integrated Gradients, Deep-Lift, Kernel SHAP, and Lime using univariate simulated time series data with temporal or latent patterns. In addition, to emphasize the severity of the latent feature saliency detection problem, we also run experiments on a real-world predictive maintenance dataset with known latent patterns. We identify Integrated Gradients, Deep-Lift, and the input-cell attention mechanism as potential candidates for refinement to yield latent saliency scores. Finally, we provide recommendations on using saliency methods for time series classification and suggest a guideline for developing latent saliency methods for time series.
The pattern graph framework solves a wide range of missing data problems with nonignorable mechanisms. However, it faces two challenges of assessability and interpretability, particularly important in safety-critical problems such as clinical diagnosis: (i) How can one assess the validity of the framework's a priori assumption and make necessary adjustments to accommodate known information about the problem? (ii) How can one interpret the process of exponential tilting used for sensitivity analysis in the pattern graph framework and choose the tilt perturbations based on meaningful real-world quantities? In this paper, we introduce Informed Sensitivity Analysis, an extension of the pattern graph framework that enables us to incorporate substantive knowledge about the missingness mechanism into the pattern graph framework. Our extension allows us to examine the validity of assumptions underlying pattern graphs and interpret sensitivity analysis results in terms of realistic problem characteristics. We apply our method to a prevalent nonignorable missing data scenario in clinical research. We validate and compare our method's results of our method with a number of widely-used missing data methods, including Unweighted CCA, KNN Imputer, MICE, and MissForest. The validation is done using both boot-strapped simulated experiments as well as real-world clinical observations in the MIMIC-III public dataset.
Despite the extensive literature on missing data theory and cautionary articles emphasizing the importance of realistic analysis for healthcare data, a critical gap persists in incorporating domain knowledge into the missing data methods. In this paper, we argue that the remedy is to identify the key scenarios that lead to data missingness and investigate their theoretical implications. Based on this proposal, we first introduce an analysis framework where we investigate how different observation agents, such as physicians, influence the data availability and then scrutinize each scenario with respect to the steps in the missing data analysis. We apply this framework to the case study of observational data in healthcare facilities. We identify ten fundamental missingness scenarios and show how they influence the identification step for missing data graphical models, inverse probability weighting estimation, and exponential tilting sensitivity analysis. To emphasize how domain-informed analysis can improve method reliability, we conduct simulation studies under the influence of various missingness scenarios. We compare the results of three common methods in medical data analysis: complete-case analysis, Missforest imputation, and inverse probability weighting estimation. The experiments are conducted for two objectives: variable mean estimation and classification accuracy. We advocate for our analysis approach as a reference for the observational health data analysis. Beyond that, we also posit that the proposed analysis framework is applicable to other medical domains.
The main intervention for coronary artery disease is stent implantation. We aim to predict post-intervention target lesion failure (TLF) months before its onset, an extremely challenging task in clinics. This post-intervention decision support tool helps physicians to identify at-risk patients much earlier and to inform their follow-up care. We developed a novel machine-learning model with three components: a TLF predictor at discharge via a combination of nine conventional models and a super-learner, a risk score predictor for time-to-TLF, and an update function to manage the size of the at-risk cohort. We collected data in a prospective study from 120 medical centers in over 25 countries. All 1975 patients were enrolled during Phase I (2016–2020) and were followed up for five years post-intervention. During Phase I, 151 patients (7.6%) developed TLF, which we used for training. Additionally, 12 patients developed TLF after Phase I (right-censored). Our algorithm successfully classifies 1635 patients as not at risk (TNR = 90.23%) and predicts TLF for 86 patients (TPR = 52.76%), outperforming its training by identifying 33% of the right-censored patients. We also compare our model against five state of the art models, outperforming them all. Our prediction tool is able to optimize for both achieving higher sensitivity and maintaining a reasonable size for the at-risk cohort over time.
Despite the extensive literature on missing data theory and cautionary articles emphasizing the importance of realistic analysis for healthcare data, a critical gap persists in incorporating domain knowledge into missing data problem formulation, assumption specification, and method development. In this paper, we highlight the gap particularly for observational data from healthcare facilities. We address this gap by identifying ten fundamental missingness scenarios arising during data measurement, recording, and pre-processing in observational health data, influenced by physicians, patients, healthcare facilities, and data scientists. We analyze the effect of scenarios on estimand formulation, missing data identification, estimation, and sensitivity analysis. To emphasize how domain-informed analysis can improve method reliability, we conduct simulation studies under the influence of various missingness scenarios. We compare the results of three common methods in medical data analysis (complete-case analysis, Missforest imputation, and inverse probability weighting estimation) for two estimands (variable mean estimation and classification accuracy). We advocate for our analysis approach as a reference for the analysis of observational health data. Furthermore, we posit that the proposed analysis framework is applicable to other medical domains, including medical wearable data analysis.