The impact of feedback training on prediction of cancer clinical trial results.

INTRODUCTION: Funders must make difficult decisions about which squared treatments to prioritize for randomized trials. Earlier research suggests that experts have no ability to predict which treatments will vindicate their promise. We tested whether a brief training module could improve experts' trial predictions. METHODS: We randomized a sample of breast cancer and hematology-oncology experts to the presence or absence of a feedback training module where experts predicted outcomes for five recently completed randomized controlled trials and received feedback on accuracy. Experts then predicted primary outcome attainment for a sample of ongoing randomized controlled trials. Prediction skill was assessed by Brier scores, which measure the average deviation between their predictions and actual outcomes. Secondary outcomes were discrimination (ability to distinguish between positive and non-positive trials) and calibration (higher predictions reflecting higher probability of trials being positive). RESULTS: A total of 148 experts (46 for breast cancer, 54 for leukemia, and 48 for lymphoma) were randomized between May and December 2017 and included in the analysis (1217 forecasts for 25 trials). Feedback did not improve prediction skill (mean Brier score for control: 0.22, 95% confidence interval = 0.20-0.24 vs feedback arm: 0.21, 95% confidence interval = 0.20-0.23; p = 0.51). Control and feedback arms showed similar discrimination (area under the curve = 0.70 vs 0.73, p = 0.24) and calibration (calibration index = 0.01 vs 0.01, p = 0.81). However, experts in both arms offered predictions that were significantly more accurate than uninformative forecasts of 50% (Brier score = 0.25). DISCUSSION: A short training module did not improve predictions for cancer trial results. However, expert communities showed unexpected ability to anticipate positive trials.Pre-registration record: https://aspredicted.org/4ka6r.pdf.

Predicting Clinical Trial Results: A Synthesis of Five Empirical Studies and Their Implications.

Expectations about future events underlie practically every decision we make, including those in medical research. This paper reviews five studies undertaken to assess how well medical experts could predict the outcomes of clinical trials. It explains why expert trial forecasting was the focus of study and argues that forecasting skill affords insights into the quality of expert judgment and might be harnessed to improve decision-making in care, policy, and research. The paper also addresses potential criticisms of the research agenda and summarizes key findings from the five studies of trial forecasting. Together, the studies suggest that trials frequently deliver surprising results to expert communities and that individual experts are often uninformative when it comes to forecasting trial outcome and recruitment. However, the findings also suggest that expert forecasts often contain a "signal" about whether a trial will be positive, especially when forecasts are aggregated. The paper concludes with needs for further research and tentative policy recommendations.

Principal investigators over-optimistically forecast scientific and operational outcomes for clinical trials.

OBJECTIVE: To assess the accuracy of principal investigators' (PIs) predictions about three events for their own clinical trials: positivity on trial primary outcomes, successful recruitment and timely trial completion. STUDY DESIGN AND SETTING: A short, electronic survey was used to elicit subjective probabilities within seven months of trial registration. When trial results became available, prediction skill was calculated using Brier scores (BS) and compared against uninformative prediction (i.e. predicting 50% all of the time). RESULTS: 740 PIs returned surveys (16.7% response rate). Predictions on all three events tended to exceed observed event frequency. Averaged PI skill did not surpass uninformative predictions (e.g., BS = 0.25) for primary outcomes (BS = 0.25, 95% CI 0.20, 0.30) and were significantly worse for recruitment and timeline predictions (BS 0.38, 95% CI 0.33, 0.42; BS = 0.52, 95% CI 0.50, 0.55, respectively). PIs showed poor calibration for primary outcome, recruitment, and timelines (calibration index = 0.064, 0.150 and 0.406, respectively), modest discrimination in primary outcome predictions (AUC = 0.76, 95% CI 0.65, 0.85) but minimal discrimination in the other two outcomes (AUC = 0.64, 95% CI 0.57, 0.70; and 0.55, 95% CI 0.47, 0.62, respectively). CONCLUSION: PIs showed overconfidence in favorable outcomes and exhibited limited skill in predicting scientific or operational outcomes for their own trials. They nevertheless showed modest ability to discriminate between positive and non-positive trial outcomes. Low survey response rates may limit generalizability.

Can Oncologists Predict the Efficacy of Treatments in Randomized Trials?

BACKGROUND: Decisions about trial funding, ethical approval, or clinical practice guideline recommendations require expert judgments about the potential efficacy of new treatments. We tested whether individual and aggregated expert opinion of oncologists could predict reliably the efficacy of cancer treatments tested in randomized controlled trials. MATERIALS AND METHODS: An international sample of 137 oncologists specializing in genitourinary, lung, and colorectal cancer provided forecasts on primary outcome attainment for five active randomized cancer trials within their subspecialty; skill was assessed using Brier scores (BS), which measure the average squared deviation between forecasts and outcomes. RESULTS: A total of 40% of trials in our sample reported positive primary outcomes. Experts generally anticipated this overall frequency (mean forecast, 34%). Individual experts on average outperformed random predictions (mean BS = 0.29 [95% confidence interval (CI), 0.28-0.33] vs. 0.33) but underperformed prediction algorithms that always guessed 50% (BS = 0.25) or that were trained on base rates (BS = 0.19). Aggregating forecasts improved accuracy (BS = 0.25; 95% CI, 0.16-0.36]). Neither individual experts nor aggregated predictions showed appreciable discrimination between positive and nonpositive trials (area under the curve of a receiver operating characteristic curve, 0.52 and 0.43, respectively). CONCLUSION: These findings are based on a limited sample of trials. However, they reinforce the importance of basing research and policy decisions on the results of randomized trials rather than expert opinion or low-level evidence. IMPLICATIONS FOR PRACTICE: Predictions of oncologists, either individually or in the aggregate, did not anticipate reliably outcomes for randomized trials in cancer. These findings suggest that pooled expert opinion about treatment efficacy is no substitute for randomized trials. They also underscore the challenges of using expert opinion to prioritize interventions for clinical trials or to make recommendations in clinical practice guidelines.

Comparison of Patient and Expert Perceptions of the Attainment of Research Milestones in Parkinson's Disease.

BACKGROUND: Commentators suggest that patients have unrealistic expectations about the pace of research advances and that such expectations interfere with patient decision-making. OBJECTIVE: The objective of this study was to compare expert expectations about the timing of research milestone attainment with those of patients who follow Parkinson's disease (PD) research. METHODS: Patients with PD and experts were asked to provide forecasts about 11 milestones in PD research in an online survey. PD experts were identified from a Michael J. Fox Foundation database, highly ranked neurology centers in the United States and Canada, and corresponding authors of articles on PD in top medical journals. Patients with PD were recruited through the Michael J. Fox Foundation. We tested whether patient forecasts differed on average from expert forecasts. We also tested whether differences between patient forecasts and the average expert forecasts were associated with any demographic factors. RESULTS: A total of 256 patients and 249 PD experts completed the survey. For 9 of the 11 milestones, patients' forecasts were on average higher than those of experts. Only exercise therapy met our 10% difference threshold for practical significance. Education was the only demographic that predicted patient deviations from expert forecasts on milestone forecasts. Patients offered significantly higher forecasts than experts that the clinical trials used in milestone queries would report positive primary outcomes. CONCLUSIONS: Differences between patient and expert expectations about research milestones were generally minor, suggesting that there is little cause for concern that patients who follow PD research are unduly swayed by inaccurate representations of research advancement in the media or elsewhere. © 2020 The Authors. Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.

Quantifying machine influence over human forecasters.

Crowdsourcing human forecasts and machine learning models each show promise in predicting future geopolitical outcomes. Crowdsourcing increases accuracy by pooling knowledge, which mitigates individual errors. On the other hand, advances in machine learning have led to machine models that increase accuracy due to their ability to parameterize and adapt to changing environments. To capitalize on the unique advantages of each method, recent efforts have shown improvements by "hybridizing" forecasts-pairing human forecasters with machine models. This study analyzes the effectiveness of such a hybrid system. In a perfect world, independent reasoning by the forecasters combined with the analytic capabilities of the machine models should complement each other to arrive at an ultimately more accurate forecast. However, well-documented biases describe how humans often mistrust and under-utilize such models in their forecasts. In this work, we present a model that can be used to estimate the trust that humans assign to a machine. We use forecasts made in the absence of machine models as prior beliefs to quantify the weights placed on the models. Our model can be used to uncover other aspects of forecasters' decision-making processes. We find that forecasters trust the model rarely, in a pattern that suggests they treat machine models similarly to expert advisors, but only the best forecasters trust the models when they can be expected to perform well. We also find that forecasters tend to choose models that conform to their prior beliefs as opposed to anchoring on the model forecast. Our results suggest machine models can improve the judgment of a human pool but highlight the importance of accounting for trust and cognitive biases involved in the human judgment process.

Wisdom of the expert crowd prediction of response for 3 neurology randomized trials.

OBJECTIVE: To explore the accuracy of combined neurology expert forecasts in predicting primary endpoints for trials. METHODS: We identified one major randomized trial each in stroke, multiple sclerosis (MS), and amyotrophic lateral sclerosis (ALS) that was closing within 6 months. After recruiting a sample of neurology experts for each disease, we elicited forecasts for the primary endpoint outcomes in the trial placebo and treatment arms. Our main outcome was the accuracy of averaged predictions, measured using ordered Brier scores. Scores were compared against an algorithm that offered noncommittal predictions. RESULTS: Seventy-one neurology experts participated. Combined forecasts of experts were less accurate than a noncommittal prediction algorithm for the stroke trial (pooled Brier score = 0.340, 95% subjective probability interval [sPI] 0.340 to 0.340 vs 0.185 for the uninformed prediction), and approximately as accurate for the MS study (pooled Brier score = 0.107, 95% confidence interval [CI] 0.081 to 0.133 vs 0.098 for the noncommittal prediction) and the ALS study (pooled Brier score = 0.090, 95% CI 0.081 to 0.185 vs 0.090). The 95% sPIs of individual predictions contained actual trial outcomes among 44% of experts. Only 18% showed prediction skill exceeding the noncommittal prediction. Independent experts and coinvestigators achieved similar levels of accuracy. CONCLUSION: In this first-of-kind exploratory study, averaged expert judgments rarely outperformed noncommittal forecasts. However, experts at least anticipated the possibility of effects observed in trials. Our findings, if replicated in different trial samples, caution against the reliance on simple approaches for combining expert opinion in making research and policy decisions.

Forecasts for the Attainment of Major Research Milestones in Parkinson's Disease.

BACKGROUND: Projections about when research milestones will be attained are often of interest to patients and can help inform decisions about research funding and health system planning. OBJECTIVE: To collect aggregated expert forecasts on the attainment of 11 major research milestones in Parkinson's disease (PD). METHODS: Experts were asked to provide predictions about the attainment of 11 milestones in PD research in an online survey. PD experts were identified from: 1) The Michael J. Fox Foundation for Parkinson's Research data base, 2) doctors specializing in PD at top ranked neurology centers in the US and Canada, and 3) corresponding authors of articles on PD in top medical journals. Judgments were aggregated using coherence weighting. We tested the relationship between demographic variables and individual judgments using a linear regression. RESULTS: 249 PD experts completed the survey. In the aggregate, experts believed that new treatments like gene therapy for monogenic PD, immunotherapy and cell therapy had 56.1%, 59.7%, and 66.6% probability, respectively of progressing in the clinical approval process within the next 10 years. Milestones involving existing management approaches, like the approval of a deep brain stimulation device or a body worn sensor had 78.4% and 82.2% probability of occurring within the next 10 years. Demographic factors were unable to explain deviations from the aggregate forecast (R2 = 0.029). CONCLUSIONS: Aggregated expert opinion suggests that milestones for the advancement of new treatment options for PD are still many years away. However, other improvements in PD diagnosis and management are believed to be near at hand.

The Role of Type and Source of Uncertainty on the Processing of Climate Models Projections.

Scientists agree that the climate is changing due to human activities, but there is less agreement about the specific consequences and their timeline. Disagreement among climate projections is attributable to the complexity of climate models that differ in their structure, parameters, initial conditions, etc. We examine how different sources of uncertainty affect people's interpretation of, and reaction to, information about climate change by presenting participants forecasts from multiple experts. Participants viewed three types of sets of sea-level rise projections: (1) precise, but conflicting; (2) imprecise, but agreeing, and (3) hybrid that were both conflicting and imprecise. They estimated the most likely sea-level rise, provided a range of possible values and rated the sets on several features - ambiguity, credibility, completeness, etc. In Study 1, everyone saw the same hybrid set. We found that participants were sensitive to uncertainty between sources, but not to uncertainty about which model was used. The impacts of conflict and imprecision were combined for estimation tasks and compromised for feature ratings. Estimates were closer to the experts' original projections, and sets were rated more favorably under imprecision. Estimates were least consistent with (narrower than) the experts in the hybrid condition, but participants rated the conflicting set least favorably. In Study 2, we investigated the hybrid case in more detail by creating several distinct interval sets that combine conflict and imprecision. Two factors drive perceptual differences: overlap - the structure of the forecast set (whether intersecting, nested, tangent, or disjoint) - and asymmetry - the balance of the set. Estimates were primarily driven by asymmetry, and preferences were primarily driven by overlap. Asymmetric sets were least consistent with the experts: estimated ranges were narrower, and estimates of the most likely value were shifted further below the set mean. Intersecting and nested sets were rated similarly to imprecision, and ratings of disjoint and tangent sets were rated like conflict. Our goal was to determine which underlying factors of information sets drive perceptions of uncertainty in consistent, predictable ways. The two studies lead us to conclude that perceptions of agreement require intersection and balance, and overly precise forecasts lead to greater perceptions of disagreement and a greater likelihood of the public discrediting and misinterpreting information.