Human judgment error in the intensive care unit: a perspective on bias and noise

Abstract

Background

In the Intensive Care Unit (ICU), clinicians frequently make complex, high-stakes judgments, where inaccuracies can profoundly affect patient outcomes. This perspective examines human judgment error in ICU settings, specifically bias (systematic error) and noise (random error). While past research has emphasized bias, we explore the role of noise in clinical decision making and its mitigation.

Main body

System noise refers to unwanted variability in judgments that should ideally be identical. This variability stems from level noise (variability in clinicians’ average judgments), stable pattern noise (variability in clinicians’ responses to specific patient characteristics), and occasion noise (random, within-clinician variability). Two strategies to reduce noise are the use of algorithms and the averaging of independent judgments.

Conclusion

Recognizing and addressing noise in clinical decision making is essential to enhancing judgment accuracy in critical care. By implementing effective noise reduction strategies, clinicians can reduce errors and improve patient outcomes, ultimately advancing the quality of care delivered in ICU settings.

Introduction

In the intensive care unit (ICU), clinicians frequently make high-stakes decisions that rely heavily on human judgment. Any errors in judgment can have profound consequences for patient outcomes, underscoring the need to minimize them. Judgment error, defined as the difference between a clinician’s prediction and the actual clinical outcome, consists of two distinct components: bias and noise [1].

Bias refers to systematic error, often caused by cognitive biases [2], such as anchoring or availability bias [3], but also includes discriminatory biases against various social groups [4]. For instance, pain assessments may be consistently underestimated for taller individuals compared to shorter ones. However, when specific biases are not shared equally among clinicians, they contribute to random variability or noise. Noise represents random error [1]. For example, noise occurs when different clinicians make varying judgments for the same patient, a phenomenon widely observed across various medical fields [5,6,7].

Recent medical research has brought critical attention to cognitive biases in human judgment [2, 8] and discriminatory biases in artificial intelligence (AI) models [9,10,11]. However, improving judgment accuracy in the ICU requires addressing not only bias—whether in humans or AI—but also the often-overlooked issue of human judgment noise. This is crucial because reducing either noise or bias by the same magnitude results in an equivalent reduction in overall error, although the strategies for mitigating them differ [1]. Accordingly, this perspective explores different types of human judgment noise, as shown in Fig. 1, and proposes two strategies for its reduction: using algorithms and averaging independent judgments.

Fig. 1

Schematic Presentation of Judgment Error

Types of human judgment noise

System noise. The ICU operates as a system designed to manage a diverse range of patients, implicitly assuming that clinicians are interchangeable in their decision making. However, this assumption is often undermined by disagreements among clinicians evaluating the same patients. This is called system noise: unwanted variability in judgments that should ideally be identical [1]. For example, in a recent pilot study,¹ 20 ICU clinicians were asked to estimate the probability of mortality (on a scale of 0% to 100%) for the same patient cases. As the second column of Table 1 demonstrates, there is variability in their estimates for the same patients, that is, system noise. System noise can be further divided into two components that explain this variability: level noise and pattern noise.

Table 1 Mortality probability estimates (%) illustrating different types of noise

Level noise. Level noise refers to the variability in the average judgment of individuals [1]. For example, some clinicians may consistently favor surgical intervention, while others may be less inclined, regardless of patient specifics. Level noise is also evident in the last column of Table 1; clinicians differ in their average mortality estimates across cases. These differences in individual judgment averages contribute to system noise. However, they do not fully explain the variability in assessments for specific patients.

Stable pattern noise. Pattern noise refers to variability between clinicians that goes beyond level noise, arising from their differing responses to specific patient characteristics. Pattern noise can be divided into stable pattern noise and occasion noise. Stable pattern noise reflects stable patterns or interaction effects between a clinician and patient traits [1]. For instance, Clinician A might consistently assign higher mortality probabilities to patients with specific comorbidities compared to both their own average estimate and those of colleagues for similar cases.² This variability in judgments does not stem from clinicians being more optimistic or pessimistic overall, but from the distinct ways in which they weigh different patient factors [4].

Occasion noise. While level noise and stable pattern noise reflect systematic between-clinician variability, occasion noise captures random, within-clinician variability. Occasion noise occurs when the same clinician makes different judgments about identical cases depending on the occasion. Common sources of occasion noise include mood, weather, fatigue, and time of day [1]. Additionally, the second-to-last column of Table 1 displays occasion noise: when patient case 3 was unknowingly repeated as case 51, inconsistencies emerged within the clinicians' own judgments.

In summary, system noise stems from three different sources: level noise, stable pattern noise, and occasion noise. This variability undermines the reliability of clinical decision making, as this process should ideally remain consistent regardless of the clinician or random situational factors. The presence of system noise means that similarly situated patients may receive different mortality predictions and subsequent treatment simply based on which clinician assesses them, introducing an element of a "lottery" effect in the system [4]. Without deliberate strategies to mitigate human judgment noise, these random errors persist, compounding across individual patients rather than canceling each other out, ultimately impacting patient care and outcomes.

Reducing human judgment noise

Bias and noise together contribute to judgment error [1]. In clinical practice, recognizing these errors and knowing how to reduce them is essential for professional health care. In the ICU context, two potential strategies for reducing noise are: using algorithms (including AI models) and averaging independent judgments.

Algorithms. A wide range of judgments and decisions are made in the ICU, and estimating mortality is one example that highlights the challenges involved. Estimating mortality is particularly complex due to the highly variable nature of ICU patients. This variability poses significant challenges to the development of intuitive expertise, which relies on several key conditions: clear, timely feedback, stable relationships between predictors and outcomes, and the opportunity to learn these relationships [12]. These requirements may not always be fulfilled in the ICU considering, for example, the results of Cox et al. [13] who found that experienced physicians were not significantly more accurate than medical students in predicting ICU mortality. Especially in these situations, algorithms offer a promising solution for improving judgment accuracy [14,15,16,17].

A major advantage of algorithms is that they reduce noise in several ways. First, well-designed algorithms apply appropriate weights to predictive variables and ignore those without predictive validity. In contrast, humans generally struggle with effectively ignoring irrelevant variables [18, 19], contributing to noise. Secondly, and potentially more importantly, algorithms consistently apply the same weighting of variables across all cases. This is not true for human judges, who vary their weighting across cases, often in ways that introduce substantial noise rather than enhancing accuracy [20]. The importance of consistency is highlighted by research showing that linear models representing experts’ judgment policies outperform the experts they were derived from [20] and that even “improper” models—such as those with unit or random weights—can outperform experts due to their consistency [21, 22].

Translating this to mortality estimates in the ICU, the Acute Physiology and Chronic Health Evaluation (APACHE) score may serve as an illustrative algorithm [23, 24]. The APACHE scoring system standardizes data collection and combination to assess patients' disease severity and mortality probability. It ensures that clinicians (a) collect the same valid variables, and (b) use a consistent algorithm to combine this data. Thereby the APACHE score enables multiple clinicians to generate the same score for identical patients, minimizing system noise (except for potential inaccuracies in the measurement instruments or inconsistencies in judgment-based variables, such as the Glasgow Coma Scale) and improving judgment accuracy compared to non-standardized assessments.

This expectation is supported by data from Cox et al. [13]. In their field study of 827 patients, Cox et al. found that physicians' predictions of in-hospital mortality had an Area Under the Curve (AUC) of 0.68 (95% CI 0.63–0.73). In comparison, our analysis of the same dataset yielded an AUC of 0.83 (95% CI 0.79–0.88) for the APACHE IV score, a significant improvement. While acknowledging an imbalance in available information between physicians and the APACHE IV, the key takeaway is that algorithmic approaches eliminate human judgment noise, underscoring the advantage of such methods.

Averaging Independent Judgments. A second method for reducing noise is averaging independent or inter-individual judgments, commonly referred to as the “wisdom of the crowd” principle [25]. This approach works because, statistically, averaging independent judgments reduces noise by the square root of the number of judgments averaged. The underlying reason for this is that independent judgments are subject to random errors, but these errors tend to cancel each other out when averaged, leading to a more accurate overall estimate. However, it is important to note that while averaging reduces noise, it does not address bias [1]. Notably, combining intra-individual judgments—such as averaging multiple judgments made by the same clinician over time—can also enhance accuracy, but only if these judgments are non-redundant, meaning they are based on different perspectives or pieces of information [26].

Discussion

This perspective focused on human judgment error, discussing two sources of error: bias and noise. We focused on noise and explained that system noise (i.e., unwanted variability in judgments for identical patients) stems from level, stable pattern, and occasion noise. Additionally, we discussed two ways—using algorithms and averaging independent judgments—to reduce noise.

While we are optimistic about the potential of algorithms, several issues need to be addressed to ensure that their implementation improves patient outcomes. First, the performance of algorithms must be critically examined, as their effectiveness cannot be assumed based on training data performance alone. Providing clinicians with outputs from incorrect models can decrease their judgment accuracy [27]. Moreover, training models based on noisy, human-labeled data—such as diagnostic labels derived from inconsistent expert judgments—can introduce errors into the models [28]. Beyond human error, inherent algorithmic challenges, such as hallucinations, inconsistencies, faithlessness, poor handling of data scarcity, and difficulty managing complexity, can undermine their performance. Secondly, algorithms can have unintended consequences, such as reinforcing social inequalities if they are discriminatory [4, 9,10,11]. These risks highlight the importance of responsible algorithm use to ensure they achieve their intended clinical benefits.

Clinicians will not mindlessly apply an algorithm’s output without additional interpretation. Clinicians synthesize an algorithm’s output as one cue within their broader judgment process, a process known as holistic synthesis [29, 30]. Holistic synthesis is often seen as advantageous, as clinicians may notice factors not incorporated into the model [17, 31]. However, whether it is indeed advantageous requires investigation, as an important and counterintuitive finding from research in other fields is that human judges often identify too many exceptions to algorithms, introducing noise and ultimately reducing accuracy [17, 30, 32,33,34,35]. Future research could explore whether presenting an algorithm’s output as one of several cues for holistic synthesis enhances or reduces judgmental accuracy compared to its unaltered application. If holistic synthesis proves beneficial, identifying which aspects of clinicians’ judgments add value could inform model refinement, ensuring these elements are reliably incorporated.

Beyond synthesizing an algorithm’s output with clinical judgment, sometimes an algorithm's output is ignored [36]. A contributing factor to this underutilization is the lack of transparency in underlying models or the definitions of predictors [36,37,38], which causes distrust in the validity of the algorithm. This is understandable; however, it is important to remember that human decision making is also a ‘black box.’ Research has shown that it is often difficult for individuals, including clinicians, to articulate their decision-making processes [39]. This makes improving mental judgments challenging. Despite the opacity of some algorithms, they can still be rigorously examined and their behavior experimentally tested—something far more difficult to achieve with human decision-makers [40].

We also recognize that making completely independent judgments is not always feasible in real-world ICU settings. However, it is crucial to recognize that independence plays a fundamental role in reducing noise, as it helps to minimize the influence of individual biases and idiosyncratic tendencies. When independence is compromised, for instance, through social influence or group dynamics, the effectiveness of noise reduction through averaging is substantially undermined [1]. To address this, systems could be designed to enable clinicians to make independent assessments and systematically aggregate them.

For critical care clinicians, recognizing judgment errors as a combination of bias and noise is crucial. This understanding underscores the importance of adopting evidence-based interventions to minimize errors. By integrating these strategies into ICU protocols, clinicians can improve the accuracy of high-stakes judgments and decisions, ultimately enhancing patient outcomes and advancing the quality of care in ICU settings.