Fast And Frugal Trees Research Articles

Identifying trigger cues for hospital blood transfusions based on ensemble of machine learning methods

BackgroundTraumatic shock is the leading cause of preventable death with most patients dying within the first six hours from arriving to the hospital. This underscores the importance of prehospital interventions, and growing evidence suggests prehospital transfusion improves survival. Optimizing transfusion triggers in the prehospital setting is key to improving outcomes for patients in hemorrhagic shock. Our objective was to identify factors associated with early in-hospital transfusion requirements available to prehospital clinicians in the field to develop a simple algorithm for prehospital transfusion, particularly for patients with occult shock.MethodsWe included trauma patients transported by a single critical care transport service to a level I trauma center between 2012 and 2019. We used logistic regression, Fast and Frugal Trees (FFTs), and Bayesian analysis to identify factors associated with early in-hospital blood transfusion as a potential trigger for prehospital transfusion.ResultsWe included 2,157 patients transported from the scene or emergency department (ED) of whom 207 (9.60%) required blood transfusion within four hours of admission. The mean age was 47 (IQR = 28 – 62) and 1,480 (68.6%) patients were male. From 13 clinically relevant factors for early hospital transfusions, four were incorporated into the FFT in following order: 1) SBP, 2) prehospital lactate concentration, 3) Shock Index, 4) AIS of chest (sensitivity = 0.81, specificity = 0.71). The chosen thresholds were similar to conventional ones. Using conventional thresholds resulted in lower model sensitivity. Consistently, prehospital lactate was among most decisive factors of hospital transfusions identified by Bayesian analysis (OR = 2.31; 95% CI 1.55 – 3.37).ConclusionsUsing an ensemble of frequentist statistics, Bayesian analysis and machine learning, we developed a simple, clinically relevant prehospital algorithm to help identify patients requiring transfusion within 4 h of hospital arrival.

Clinical decision making in physical therapy – Exploring the ‘heuristic’ in clinical practice

Clinical decision-making (CDM) plays an integral role in the work of a physical therapist and has ramifications for patient outcomes and experience. Rational decision-making – acting in a manner that helps us achieve our goals – is influenced by cognitive, emotional, and social variables. The dual process theory helps us understand how clinicians make what they perceive to be rational decisions. Within dual process is the use of cognitive decisional shortcuts, commonly referred to as ‘heuristics,’ which are either developed through experience or the use of fast and frugal trees (FFT). The use of heuristics in physical therapy practice has yet to be explored. This paper aims to describe this subset of physical therapy decision-making and to identify the typical cognitive biases - the error in heuristic-driven decision making - inherent in this style of reasoning. Common heuristics and their related biases are described and illustrated with vignettes, including the anchoring, availability, confirmatory, and representative heuristics. The authors conclude by proposing interventions to optimize physical therapists' awareness of their use of heuristics in clinical decision-making and to identify and minimize the influence of potential bias.

Prognostic Value of the Neurological Pupil Index in Patients With Acute Subarachnoid Hemorrhage.

The Neurological Pupil index (NPi) provides a quantitative assessment of pupil reactivity and may have prognostic value in patients with subarachnoid hemorrhage (SAH). We aimed to explore associations between the NPi and clinical outcomes in patients with SAH. A retrospective analysis of 79 consecutive patients with acute SAH. Age, sex, Acute Physiology and Chronic Health Evaluation-II score, and respiratory failure and NPi in each eye were recorded at admission. The primary outcomes included death and poor clinical outcome (defined as inpatient death, care withdrawal, or discharge Glasgow Outcome Score <4). Groups were compared using the Fisher exact test, and predictive models developed with fast-and-frugal trees (FFTs). A total of 53 patients were included: 21 (40%) had poor clinical outcomes and 2 (4%) died. Univariate analysis found that only APACHE-II score (P < 0.001) and respiratory failure (P = 0.04) were significantly associated with poor clinical outcomes. NPi was lower among patients with poor clinical outcomes (mean 4.3 in the right eye and 4.2 in the left eye) vs those without (mean 4.5 in the right eye and 4.5 in the left eye), but neither was significant. However, the most accurate FFTs for death and poor clinical outcome included NPi after accounting for age in the death FFT and APACHE-II score in the poor outcome FFT (sensitivity [sn] = 100%, specificity [sp] = 94%, and accuracy (ac) = 94% in a model for death; sn = 100%, sp = 50%, and ac = 70%) in a model for poor clinical outcome. Our study supports the NPi as a useful prognostic marker for poor outcomes in acute SAH after accounting for age and APACHE-II score.

Fluid Overload Phenotypes in Critical Illness-A Machine Learning Approach.

Background: The detrimental impact of fluid overload (FO) on intensive care unit (ICU) morbidity and mortality is well known. However, research to identify subgroups of patients particularly prone to fluid overload is scarce. The aim of this cohort study was to derive “FO phenotypes” in the critically ill by using machine learning techniques. Methods: Retrospective single center study including adult intensive care patients with a length of stay of ≥3 days and sufficient data to compute FO. Data was analyzed by multivariable logistic regression, fast and frugal trees (FFT), classification decision trees (DT), and a random forest (RF) model. Results: Out of 1772 included patients, 387 (21.8%) met the FO definition. The random forest model had the highest area under the curve (AUC) (0.84, 95% CI 0.79–0.86), followed by multivariable logistic regression (0.81, 95% CI 0.77–0.86), FFT (0.75, 95% CI 0.69–0.79) and DT (0.73, 95% CI 0.68–0.78) to predict FO. The most important predictors identified in all models were lactate and bicarbonate at admission and postsurgical ICU admission. Sepsis/septic shock was identified as a risk factor in the MV and RF analysis. Conclusion: The FO phenotypes consist of patients admitted after surgery or with sepsis/septic shock with high lactate and low bicarbonate.

Age at diagnosis, but not HPV type, is strongly associated with clinical course in recurrent respiratory papillomatosis.

BackgroundRecurrent Respiratory Papillomatosis (RRP) is a rare disease characterized by the growth of papillomas in the airway and especially the larynx. The clinical course is highly variable among individuals and there is poor understanding of the factors that drive an aggressive vs an indolent course.MethodsA convenience cohort of 339 affected subjects with papillomas positive for only HPV6 or HPV11 and clinical course data available for 1 year or more, from a large multicenter international study were included. Exploratory data analysis was conducted followed by inferential analyses with frequentist and Bayesian statistics.ResultsWe examined 339 subjects: 82% were diagnosed prior to the age of 18 years, 65% were infected with HPV6, and 69% had an aggressive clinical course. When comparing age at diagnosis with clinical course, the probability of aggressiveness is high for children under five years of age then drops rapidly. For patients diagnosed after the age of 10 years, an indolent course is more common. After accounting for confounding between HPV11 and young age, HPV type was minimally associated with aggressiveness. Fast and Frugal Trees (FFTs) were utilized to determine which algorithms yield the highest accuracy to classify patients as having an indolent or aggressive clinical course and consistently created a branch for diagnostic age at ~5 years old. There was no reliable strong association between clinical course and socioeconomic or parental factors.ConclusionIn the largest cohort of its type, we have identified a critical age at diagnosis which demarcates a more aggressive from less aggressive clinical course.

Precision medicine for lung cancer decision-making: Evaluation of an -omics based FFT approach to personalized medicine.

e20713 Background: The prevalence of next-generation sequencing and the availability of a large number of targeted therapies in the clinics, has complicated treatment decision-making in lung cancer. While national guidelines and commercial pathways offer a method to improve the oncologists’ adherence to appropriate testing and treatment modalities available, more effort is required to solidify this as a standard of care model at academic and community sites. A better understanding of the improved durable survival of targeted therapy assignment compared with non-targeted therapy outside of the clinical trial setting is needed to understand the efficacy and accuracy of precision medicine. Methods: We perform an in-depth analysis of a series of lung AD patients (n = 798) with genomic and clinical data in a recently created thoracic patient registry, who were treated at COH between 2009-2018 period. Results: 798 individuals with lung AD were identified in the Thoracic Oncology Registry who were treated or were intended to be treated at COH; 662 (83%) of the patients had genomic testing performed at the request of their treating oncologist and 460 (58%) of whom received a 1st-line targeted therapy decision (including clinical trial assignment based on bio-marker). Oncogenic alterations were detected in 653 (82%) patients with the majority presenting with EGFR (47%), who were mostly treated with erlotinib (78%). 462/653 (70%) patients had an alteration detected with an available FDA approved therapy and 90% (416/462) of the patients were appropriately matched to a targeted therapy based on the oncologist’s decision. Several decision-making algorithms were tested and fast-and-frugal trees (FFTs) proved superior with a positive predictive value (PPV) of 90% and only required two important cues in informing the decision of the type of treatment to give to the patient. Furthermore, a targeted therapy treatment decision showed a significant benefit with a median OS of 38 months as compared to 22 months in the non-targeted therapy decision-making group (p &lt; 0.00001). This was also evident in the PFS analysis where targeted therapy decision-making had a median survival of 9 months as compared with 5 months in the other groups (p &lt; 0.00001). Conclusions: FFTs are a novel tool to test the efficacy of precision medicine in a real-world setting and can provide a more streamlined method for clinical guidance and decision-making. FFTs were able to predict with 90% PPV a precision medicine decision that was correlated with improved PFS (9 vs 5 months) and OS (38 vs 22 months).

Financing agricultural drought risk through ex-ante cash transfers

Despite advances in drought early warning systems, forecast information is rarely used for triggering and financing early actions, such as cash transfer. Scaling up cash transfer pay-outs, and overcoming the barriers to actions based on forecasts, requires an understanding of costs resulting from False Alarms, and the potential benefits associated with appropriate early interventions. On this study, we evaluate the potential cost-effectiveness of cash transfer responses, comparing the relative costs of ex-ante cash transfers during the maize growing season to ex-post cash transfers after harvesting in Kenya. For that, we developed a forecast model using Fast-and Frugal Trees that unravels early warning relationships between climate variability, vegetation coverage, and maize yields at multiple lead times. Results indicate that our models correctly forecast low maize yield events 85% of the time across the districts studied, some already six months before harvesting. The models' performance improves towards the end of the growing season driven by a decrease of 29% in the probability of False Alarms. Overall, we show that timely cash transfers ex-ante to a disaster can often be more cost-effective than investing in ex-post expenditures. Our findings suggest that early response can yield significant cost savings, and can potentially increase the effectiveness of existing cash transfer systems.

Fast-and-frugal trees as noncompensatory models of performance-based personnel decisions

Employees’ performance provides the basis for many personnel decisions, and to make these decisions, managers often need to integrate information from different performance-related cues. We asked college students and experienced managers to make a series of performance-based personnel decisions and tested how well weighting-and-adding, compensatory logistic regression and lexicographic, noncompensatory fast-and-frugal trees (FFTs) could describe participants’ decision processes regarding both choices and reaction times. Results show that a significant proportion of the participants (i.e., nearly half of the college students and more than two-thirds of the experienced managers) applied FFTs to make such decisions, and that the majority of them adopted key features of FFTs adaptively in response to a manipulation of the required distributions of positive (bonus) or negative (termination) decisions. Overall, the process-oriented approach applied in our study provides insights on not only what cues managers use for performance-based personnel decisions, but also how they use these cues.

Tying up loose ends: a method for constructing and evaluating decision aids that meet blunt and sharp-end goals

We present a methodological framework for constructing and evaluating decision aids – fast and frugal trees (FFTs) – ideally suited to the front line of an organisation. Their performance can be analysed in signal detection theory, allowing for transparent selection of FFTs given managerial-level trade-offs among type I and II errors. We extend FFTs from binary classification to selection from multiple actions (FFT multiple) as well as performance analysis to organisational goal states beyond type I and II error reduction. Concepts and framework are introduced and a tutorial-style example application (threat assessment at military checkpoints) is provided. Throughout, we discuss ways to deal with missing or incomplete data and show that the performance of decision aids may be overestimated if the effectiveness of actions is not heeded. The methodology can be used to construct and evaluate decision aids in any area characterised by dichotomised cues and a one-to-many mapping between categorisation outcomes and actions.

A signal-detection analysis of fast-and-frugal trees.

Models of decision making are distinguished by those that aim for an optimal solution in a world that is precisely specified by a set of assumptions (a so-called “small world”) and those that aim for a simple but satisfactory solution in an uncertain world where the assumptions of optimization models may not be met (a so-called “large world”). Few connections have been drawn between these 2 families of models. In this study, the authors show how psychological concepts originating in the classic signal-detection theory (SDT), a small-world approach to decision making, can be used to understand the workings of a class of simple models known as fast-and-frugal trees (FFTs). Results indicate that (a) the setting of the subjective decision criterion in SDT corresponds directly to the choice of exit structure in an FFT; (b) the sensitivity of an FFT (measured in d) is reflected by the order of cues searched and the properties of cues in an FFT, including the mean and variance of cues’ individual ds, the intercue correlation, and the number of cues; and (c) compared with the ideal and the optimal sequential sampling models in SDT and a majority model with an information search component, FFTs are extremely frugal (i.e., do not search for much cue information), highly robust, and well adapted to the payoff structure of a task. These findings demonstrate the potential of theory integration in understanding the common underlying psychological structures of apparently disparate theories of cognition.

A signal-detection analysis of fast-and-frugal trees.

Models of decision making are distinguished by those that aim for an optimal solution in a world that is precisely specified by a set of assumptions (a so-called "small world") and those that aim for a simple but satisfactory solution in an uncertain world where the assumptions of optimization models may not be met (a so-called "large world"). Few connections have been drawn between these 2 families of models. In this study, the authors show how psychological concepts originating in the classic signal-detection theory (SDT), a small-world approach to decision making, can be used to understand the workings of a class of simple models known as fast-and-frugal trees (FFTs). Results indicate that (a) the setting of the subjective decision criterion in SDT corresponds directly to the choice of exit structure in an FFT; (b) the sensitivity of an FFT (measured in d') is reflected by the order of cues searched and the properties of cues in an FFT, including the mean and variance of cues' individual d's, the intercue correlation, and the number of cues; and (c) compared with the ideal and the optimal sequential sampling models in SDT and a majority model with an information search component, FFTs are extremely frugal (i.e., do not search for much cue information), highly robust, and well adapted to the payoff structure of a task. These findings demonstrate the potential of theory integration in understanding the common underlying psychological structures of apparently disparate theories of cognition.