Collaboration between a human group and artificial intelligence can improve prediction of multiple sclerosis course: a proof-of-principle study

Background: Multiple sclerosis has an extremely variable natural course. In most patients, disease starts with a relapsing-remitting (RR) phase, which proceeds to a secondary progressive (SP) form. The duration of the RR phase is hard to predict, and to date predictions on the rate of disease progression remain suboptimal. This limits the opportunity to tailor therapy on an individual patient's prognosis, in spite of the choice of several therapeutic options. Approaches to improve clinical decisions, such as collective intelligence of human groups and machine learning algorithms are widely investigated. Methods: Medical students and a machine learning algorithm predicted the course of disease on the basis of randomly chosen clinical records of patients that attended at the Multiple Sclerosis service of Sant'Andrea hospital in Rome. Results: A significant improvement of predictive ability was obtained when predictions were combined with a weight that depends on the consistence of human (or algorithm) forecasts on a given clinical record. Conclusions: In this work we present proof-of-principle that human-machine hybrid predictions yield better prognoses than machine learning algorithms or groups of humans alone. To strengthen and generalize this preliminary result, we propose a crowdsourcing initiative to collect prognoses by physicians on an expanded set of patients.


Introduction
The natural course of multiple sclerosis (MS) is extremely variable, ranging from extremely mild to very aggressive forms. Most patients experience an initial relapsing-remitting (RR) phase, in which symptoms appear and fade. Eventually, remissions fail and the disease proceeds to a secondary progressive (SP) form, leading to incremental disability. The palette of disease-modifying treatments is becoming relatively large, in principle opening the possibility to tailor the therapy to meet the specific needs of each patient. Unfortunately, the accuracy of parameters to predict the rate of disease progression remains suboptimal.
Being all the above therapies preventive, in the absence of exact prognostic indicators we have to accept that a proportion of patients is either under-or over-treated. This is a serious concern as the disease can be severely disabling, and some of the available therapies can lead to adverse events that can be worse than the disease itself. Thus, the possibility to formulate a prognosis as exact as possible is becoming increasingly appealing.
In the clinics, as in any other fields of human knowledge, innovative approaches based on machine learning and collective reasoning methods are used in an attempt to succeed where traditional methods of forecasting failed. Machine learning algorithms catch complex relations among existing data to an extent beyond standard regression models. Good performances have been obtained for the diagnosis of Parkinson's disease and the prognosis of disease progression in amyotrophic lateral sclerosis (Dinov et al., 2016;Küffner et al., 2015). For MS, machine learning algorithms can correctly classify disease course in about 70% of cases of both clinically definite MS and of clinically isolated syndrome (Fiorini et al., 2015;Wottschel et al., 2014;Zhao et al., 2017), a good result that still requires improvement to become of clinical value.
Through collective reasoning, or collective intelligence, groups of lay people may perform as well as experts. In principle, the larger the group, the higher the prediction accuracy (see for review Ponsonby & Mattingly, 2015), which led to the development of several crowdsourcing initiatives. Possibly, the forerunner was FOLDIT study on protein folding (Cooper et al., 2010), but crowdsourcing has been exploited also for diagnostic purposes in pathologies, such as breast cancer (Candido dos Reis et al., 2015), skin cancer (King et al., 2013) or ophtalmology (Wang et al., 2016). However, when expert people are involved, even small groups can outperform the best among them, at least when a yes/no answer to well-defined diagnostic questions is requested based on radiographic/histological images, (Kurvers et al., 2016;Sonabend et al., 2017;Wolf et al., 2015). Studies with medical students show that working in pairs, either interacting while responding (Hautz et al., 2015) or aggregating responses ex post (Kämmer et al., 2017), ameliorates diagnostic ability, with further improvements when group size increases (Hautz et al., 2015;Kämmer et al., 2017), in line with the core idea of Collective intelligence. Similar results have been obtained also for prognoses on critically ill patients (Poses et al., 1990) Combination of human and machine predictions into hybrid forecasts exploits human intuitive reasoning and computer classification capabilities, potentially boosting both. Indeed, at least in the case of predicting the course of actions in American football games within the frame of prediction markets, hybrid groups performed better than either humans or computers. (Nagar & Malone, 2011). In this paper, we report the promising results of a preliminary study on the combination of predictions made by humans with those of a machine learning algorithm on the progression of multiple sclerosis in a set of patients. Both agents (humans and computers) considered clinical data typically available to neurologists during routine visits. Magnetic Resonance Imaging (MRI) data were not included as more clinical than radiological exams are routinely performed (on average 3 visits per year vs. 1 MRI). Moreover, images, acquired and analysed at specialized centres can improve the algorithm performance (Zhao et al., 2017), but in real world imaging data usually lack the standardization required for analysis, for instance in term of head position reproducibility (Weinstock-Guttman et al., 2018), and research-grade image analysis is not routinely performed. Conversely, clinical data have recently been shown to have good predictive value (Goodin et al., 2018). Machine learning and collective intelligence performed almost equally well, but their combination yielded a small, yet statistically significant, improvement in the reliability of the forecasts on disease evolution over different time periods.
These results indicate that it is worth deepening the study of human and machine clinical predictions, as well as the potentiality of hybrid predictions, for which we propose a crowdsourcing approach on a platform specifically designed for this analysis (DiagnoShare).

Amendments from Version 1
This version addresses most of the points raised by the Reviewers, to whom we are grateful for useful criticisms.
We included several references and statements to justify why we only considered clinical data, not MRI and to document the use of crowdsourcing initiatives in the diagnostic realm.
We give more details on how we built the database and on the procedure used to obtain hybrid forecasts, which represent the core of the paper.
Finally, we repeatedly state that this work represents a proof of concept of the value of hybrid forecasts, a preliminary step towards deeper studies.

REVISED
values were provided for each parameter, referring to age, time to complete a task, clinical score or presence/absence of each symptom. For each visit, we noted if the patient was in RR or SP stage after 180, 360 and 720 days, so that predictions could be compared with the true progression of disease in each patient, reported in Supplementary File: TrueOutcomes.xlsx, where 0 means "still in RR phase", 1 indicates "transitioned to SP phase". Notice that several patients reached the SP MS stage after the last visit included in the database, so that the number (percentage of entries) of "1" records is 65 (12.3%), 125 (23.7%), and 211 (40.0%) at 180, 360, and 720 days, respectively. Data potentially revealing the identity of the patients was removed from the shared database.

Ethics
Use of database for research purposes was authorized by the Ethical committee of Sapienza University (Authorization n. 4254_ 2016, dated November 2, 2016).

Classification with machine learning
Having a correctly labelled dataset (Supplementary Table 1), in which each entry is associated to the outcome, we used the Random forest supervised approach to classification (Breiman, 2001;Liaw & Weiner, 2002), using the Scikit-learn toolbox version 0.16.1.
To benchmark the performance of the trained models, we used a modified leave-one-out approach. Since data was limited (a set of 527 records), and not independent, as it had been obtained from 84 patients, with a simple random leave-one-out the training set would be composed of many correlated same-patient data. Even worse, some of the data from patients present in the training set would be used to validate the model in the benchmarking stage. As a consequence, the model would overfit the training data, misleadingly showing very good performance. Being presented with many data from the same patient, the model optimizes its ability in recognizing patients themselves, through their highly correlated clinical variables.
To avoid these problems, we used a modified leave-one-out approach, training the algorithm with the following rules: 1. We excluded all visits from one patient from the dataset 2. We built 50 training sets, each composed by 83 records, taking care to include only one clinical record (randomly chosen) for every remaining patient 3. We trained 50 Random Forest models, one for each training set.

4.
We computed the probability of the transition from RR to SP by averaging the predictions of the 50 models on all the visits of the excluded patient. Predictions consisted in scores from 0 (Extremely unlikely) to 1 (Highly probable).
We repeated the procedure for the 84 patients, obtaining an estimation of the probability of the RR to SP transition for each of the 527 clinical records. Three different prediction delays were considered, namely 180, 360 and 720 days. Results obtained are presented in Supplementary Table 1.

Human predictions
Forty-two medical students in the final two years of their course (Sapienza University, Rome Italy, based within Sant'Andrea hospital), volunteered to participate in the task. All were familiar with clinical records in general, and were instructed on the meaning of each entry present in the medical records of MS patients. This part of the study was approved by the Ethical Committee of the Department of Physiology and Pharmacology, Sapienza University on July 13, 2017.
For adequate comparison with computer predictions, students evaluated 50 medical records, collected in a questionnaire, randomly extracted from the same dataset used for machine learning and estimated the probability that the patient would progress to the SP phase within 180, 360 and 720 days. Scores were from 0 (Extremely unlikely) to 5 (Highly probable). Predictions (see Supplementary file Student_Predictions.xlsx) were analysed, using the AUC.
On average, each clinical record was evaluated by 4 of the 42 students. Predictions were less accurate than those proposed by machine learning (Table 1). Standard deviation was larger for the 180 day time point, indicating that opinions on the long-term evolution of the disease are more widely shared, although they are not more precise. To evaluate the impact of collective intelligence, we measured the performance of Pairs, considering all visits evaluated by at least two individual students, randomly selecting only 2 scores when more were available. The prognoses were averaged before computing the AUC, which showed a marked increase ( Table 1). Aggregation of all singles (Group) yielded a further small increase in the performance of the forecasting (Table 1), which almost equalled that of random forest algorithm.

Hybrid predictions
We next integrated human and computer predictions into a hybrid prediction, which combines human clinical reasoning with the classification approach of machine learning algorithms. These different "ways of reasoning" possibly lead to quite divergent predictions on individual cases, a complementarity that should be exploited taking the difference into account when creating hybrid predictions.
The simplest approach to aggregate forecast is performing a linear or weighted average of the predictions released by humans or computer. For each clinical record at a given time point (180, 360 and 720 days), the final forecast by either agent is the average of "unitary predictions" given by several individuals or decision trees. If "unitary predictions" of one agent are highly concordant, it means that the prediction is quite obvious for the agent, suggesting that it is probably correct. We therefore ranked forecasts on clinical records in order of concordance of "unitary predictions", for the two agents separately. Then, a normalized ranking was assigned, ranging from 1 for the most consistent predictions to 0 for the most scattered and ranks were squared to emphasize the contribution of the most consistent agent. The hybrid prediction score for each clinical record was then obtained by summing the two squared ranks.
Note that a linear combination of rankings resulted in a worse performance of hybrid predictions, as the information about the most consistent prediction between the two agents was be lost. A similarly degraded performance was observed when predictions were not ranked.
Since our dataset is relatively small, as is the number of students that evaluated the clinical records, we used a bootstrap procedure to evaluate the statistical significance of the improvement. The bootstrap (Efron & Tibshirani, 1994;Felsenstein, 1985) consists in random sampling of the dataset that allows the estimation of confidence intervals.
As shown in Table 1 and Figure 1, hybrid predictions yielded a small but statistically significant (P<0.001) improvement in the prediction of disease course in time. Significance was evaluated from confidence limits using standard methods (Altman & Bland, 2011). Each student worked on a questionnaire (lines labelled "questionnaire", column B.) listing 50 clinical reports (lines labelled "Clinical report N", columns B to AY) and made a prediction on the probability of RR -to-SP transition within 180, 360 and 720 days (lines labelled Prediction @ 180, 360, 720, columns B to AY) The numbering of Clinical reports is the same used in Dataset 1.  Score_180, Score_360, Score_720: Probability that the patient will transition to SP phase within 180, 360 and 720 days after the visit in which clinical record was obtained. The numbering of Clinical reports is the same used in Dataset 1.

Discussion
A number of studies have investigated the possibility to increase the appropriateness of clinical decisions through collective intelligence of human groups (for instance, Kurvers et al., 2016;Sonabend et al., 2017;Wolf et al., 2015) or machine learning algorithms. The latter approach has been used in a great variety of tasks, and its value in the medical realm is possibly overstated (Chen & Asch, 2017). However, machine learning methods performed well for prognostic predictions (Küffner et al., 2015;Zhao et al., 2017). In particular, the Random forest approach provided good predictions on ALS course (Küffner et al., 2015).
In this work we present proof-of-principle that human-machine hybrid predictions attain prognostic ability above that of machine learning algorithms and groups of humans alone.
The duration of the RR phase before its shift into progression has always been difficult to predict, and possibly the random occurrence of relapses (Bordi et al., 2013) contributes to the lack of univocal indicators. No approach, no matter how good, can yield certainty when cause-effect relations are unknown. Thus, our aim has been to obtain predictions on the probability that MS patients in the RR phase will convert to a SP form within a certain time frame. Predictions on the course of real patients were provided by medical students and a random forest algorithm. A significant improvement of predictive ability was obtained when predictions were combined in a non-linear manner, with a weight that depends on the consistence of human (or algorithm) forecasts on a given clinical record.
This result can be considered in agreement with several studies on different medical issues showing that predictor's confidence correlates very well with the correctness of the prediction (Detsky et al., 2017;Hautz et al., 2015;Kämmer et al., 2017;Kurvers et al., 2016). Indeed, the concordance of different members of a given group (students or runs of the random forest model) can be taken as indicating that the agent is "sure" of the forecast. Further work investigating the best ways to combine predictions of different agents is ongoing.
In spite of the relatively basic machine learning technique used, the small number of students involved and their limited clinical knowledge, this work suggests that hybrid predictions can be useful to improve the prognosis of MS course. A deeper study is therefore of interest, to evaluate how general this conclusion is. To recruit more and more skilled humans, we propose a crowdsourcing initiative called DiagnoShare that is being advertised among physicians.
A reliable tool to predict MS progression can be of aid to clinicians to tailor therapy to each patient, but also in clinical trials, to evaluate whether drugs modify the estimated outcome of each enrolled patient, as proposed for ALS (Küffner et al., 2015).
In the long run, it is possible that further developments in our ability to combine collective reasoning and machine predictions will have a profound impact also on the organization and management of medical care, particularly in hospital settings.

Competing interests
No competing interests were disclosed.

Grant information
CENTERS is a special project of, and is supported by, Fondazione Italiana Sclerosi Multipla. AT and AZ acknowledge funding from the "CNR Progetto di Interesse CRISIS LAB".
All funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. I think exploring how to fuse multiple expert opinions is a very interesting line of research in computer aided diagnostics. Here, the authors test how to make use of lay persons, and I would agree that there are many tasks when a (briefly trained) lay person or non-expert can contribute significantly to an analytical task.

Open Peer Review
In the application here, I would be rather critical about this approach, though. For example, the authors write "through collective reasoning, or collective intelligence, groups of lay people may perform as well as experts." I would not agree, by any means. How would a lay person without training be able to distinguish, for example, a stroke related white matter hyper-intensity from an MS lesion? Or even a large MR artifact? Averaging will reduce variance in prediction, but the individual prediction itself has to be unbiased. In other words: the layman predictor has to be correct on average. But how would they possibly be in case they have no idea about how to read these data? Moreover, the authors point out that "studies with medical students show that working in pairs ameliorates diagnostic ability". Is this because of a better discussion of the decision? With two subjects it cannot be the power of large numbers that this study relies on.
Instead of exploring how to fuse layman's decisions, I would recommend the authors to explore how to fuse decisions of different algorithms, or from neurologists of different training/seniority level, or decisions based on different sources.

Technical:
The authors describe a "leave-one patient-out" cross-validation as an Experimental setup and evaluation: innovation of their study. While this is a good approach, it is not new.
There are different classes -what is the distribution of those classes for the 84 Algorithm and training: patients? What is in the reports? Numbers? Free text? What features are input to the random forest algorithm? How many features at all? How did you train the algorithm (parameters "mtry", why 50 trees?) Without this information it is difficult to assess whether the performance of your random forest is bad (i.e., close to layman's predictions) because of an suboptimal training procedure, or because this is a hard problem indeed. problem indeed.
(Described in the section "To compare the two sets... of the most consistent agent.") I don't Fusion rule: understand what you do. How does summing a squared ranking lead to a prediction score? I assume you are using the normalized (and squared) ranking as a sort of weight? Why do you square the rankings? What happens when you use other non-linear transformations? Is there any way you illustrate the distributions so that we can follow your reasoning? How about presenting simple rules like averaging, or majority voting at least as a baseline we can compare against?

If applicable, is the statistical analysis and its interpretation appropriate? Partly
Are all the source data underlying the results available to ensure full reproducibility? No

Are the conclusions drawn adequately supported by the results? Partly
No competing interests were disclosed.

Competing Interests:
I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.
Author Response 26 Jul 2018 , Sapienza University of Rome, Italy

Francesca Grassi
First of all, thank you for taking the time to read our work and to give useful comments. We hope that our responses will clear your doubts. We list below the changes introduced in the new version, prompted by your observations. We hope that you agree with us that it is improved In the application here, I would be rather critical about this approach, though. For example, the authors write "through collective reasoning, or collective intelligence, groups of lay people may perform as well as experts." I would not agree, by any means. How would a lay person without training be able to distinguish, for example, a stroke related white matter hyper-intensity from an MS lesion? Or even a large MR artifact? Averaging will reduce variance in prediction, but the individual prediction itself has to be unbiased. In other words: the layman predictor has to be correct on average. But how would they possibly be in case they have no idea about how to read these data? ANSWER Although your point of view is quite understandable, there is a large body of literature on the topic Although your point of view is quite understandable, there is a large body of literature on the topic of collective intelligence. In the hope to overcome your skepticism on this point, we added some more references to published work on diagnostic crowdsourcing initiatives.
Moreover, the authors point out that "studies with medical students show that working in pairs ameliorates diagnostic ability". Is this because of a better discussion of the decision? With two subjects it cannot be the power of large numbers that this study relies on. ANSWER It is now better explained that the two quoted studies use different methods: real pairs in one, aggregated opinions in the other, yet both obtain better performances. Authors do not discuss the underlying processes, so we cannot indicate the real reason of a better performance.
Instead of exploring how to fuse layman's decisions, I would recommend the authors to explore how to fuse decisions of different algorithms, or from neurologists of different training/seniority . level, or decisions based on different sources ANSWER Thank you for your suggestion. Understanding that we have to deepen out study (now repeatedly stated throughout the paper) we have developed DiagnoShare to obtain the predictions of clinicians of different expertise and we are investigating the performance of algorithm combinations.

Experimental setup and evaluation: The authors describe a "leave-one patient-out" cross-validation as an innovation of their study. While this is a good approach, it is not new.
ANSWER Thank you for your observation. Indeed, it is better to define our approach as a modified leave-one-out method. It is modified, because we not only left one patient out, we also included only one record for each of the remaining patients.
Algorithm and training: There are different classes -what is the distribution of those classes for the 84 patients? What is in the reports? Numbers? Free text? What features are input to the random forest algorithm? How many features at all? How did you train the algorithm (parameters "mtry", why 50 trees?) Without this information it is difficult to assess whether the performance of your random forest is bad (i.e., close to layman's predictions) because of an suboptimal training procedure, or because this is a hard problem indeed ANSWER Thank you for pointing out that this part of the paper required clarification. As now better emphasized in the text, in this proof-of-concept work we considered only patients that actually transitioned to the SP phase, so there is a unique class of patients. Features input to the RF algorithm are listed in Supplementary Table 1. We added a statement to declare what types of numerical values we used in the work. The results presented show that the RF algorithm performs better than layman, as its performance is however better than that of individual medical students, that are not quite laymen, although not experts as well. In any case, the focus of the paper is not on the goodness of the algorithm, but on the value of combining different approaches to the prediction problem, which indeed has been resisting solution for many years of medical analysis. We agree with you that, indeed, this point is complex and we try a different explanation, hoping that it is clearer. First of all, ranking is inherent to building a ROC curve. Since we have only two agents (humans and RF algorithm), we cannot use a majority rule, we can only perform an average (linear or weighted) of the scores. For any clinical record, the final forecast is the average of "unitary predictions" by multiple individuals or decision trees. If "unitary predictions" of one agent are highly concordant, it means that the prediction is quite obvious for the agent, suggesting that is more probably correct than others. We ranked forecasts on clinical records in order of concordance of "unitary predictions" and emphasized the value of agreement by squaring the ranks. In line with other pieces of research, this weighted average performed better than linear averaging, as stated in the paper This is an interesting and clearly written article on using a machine learning method (random forests) and medical students to form "hybrid" predictions of disease progression in MS, specifically the conversion from RRMS to SPMS. The article claims that the results are a proof-of-principle that combining machine learning and human predictions is better than either approach alone.
The main strengths of the article are its clear writing, the reproducbility of the experiments, the clinical importance of the application, and topical nature of the subject, as machine learning for clinical prediction is such a hot topic that integration with the clinical workflow is a critical area of study.
The main limitations of the article are that only clinical parameters were used to perform the predictions, and the longitudinal nature of the data was not used to its full benefit. To realize the potential of machine learning for MS prediction, imaging parameters should be included (there is good literature on MS prediction using imaging), and examining changes over time is important for both machine (eg, using recurrent networks) and human raters (examining clinical changes over multiple time points). The article places some importance on having the computer and humans using the same set of input parameters, but I do not feel that this is warranted; the data should be selected to be most appropriate for each approach.
Given the above limitations, it is difficult to generalize the findings to say that hybrid predictions are better than either machine learning or humans. This could be true, and the article provides some support for that, but more work needs to be done to provide strong evidence.

Is the study design appropriate and is the work technically sound? Partly
Are sufficient details of methods and analysis provided to allow replication by others? Yes

If applicable, is the statistical analysis and its interpretation appropriate? Yes
Are all the source data underlying the results available to ensure full reproducibility? Yes

Are the conclusions drawn adequately supported by the results? Partly
No competing interests were disclosed.

Competing Interests:
I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.
Author Response 26 Jul 2018 , Sapienza University of Rome, Italy

Francesca Grassi
First of all, thank you for taking the time to read our work and to give useful comments. We hope that our responses will clear your doubts. We list below the changes introduced in the new version, prompted by your observations. We hope that you agree with us that it is improved.
The main limitations of the article are that only clinical parameters were used to perform the predictions, and the longitudinal nature of the data was not used to its full benefit. To realize the potential of machine learning for MS prediction, imaging parameters should be included (there is good literature on MS prediction using imaging), and examining changes over time is important for both machine (eg, using recurrent networks) and human raters (examining clinical changes over multiple time points). The article places some importance on having the computer and humans using the same set of input parameters, but I do not feel that this is warranted; the data should be selected to be most appropriate for each approach. ANSWER We agree with you that many other approaches could be used. As now stated in the text, we chose to explore predictions done using only clinical data, available to all neurologists, which have recently been independently demonstrated to have good predictive value (Goodin et al., 2018; reference added to the paper). Imaging data performed in real-world clinical settings do not have the standardization required for predictions either by experts or algorithms. However, future studies aimed at confirming this proof-of-principle, initial work will surely consider different options.
Given the above limitations, it is difficult to generalize the findings to say that hybrid predictions are better than either machine learning or humans. This could be true, and the article provides some support for that, but more work needs to be done to provide strong evidence. ANSWER In this proof-of-concept study, we chose to work with medical students instead of neurologists because we wanted to test if even a group of relatively uneducated people can enhance the predictive ability of machine learning algorithms, which is now well established.
We agree with you that the next step is to obtain predictions by neurologists and other medical doctors, and in fact we set up the platform DiagnoShare (http://www.phys.uniroma1.it/diagnoshare ) to extend the study.
Hopefully, we can soon extend this work with a final study None Competing Interests: The benefits of publishing with F1000Research: Your article is published within days, with no editorial bias You can publish traditional articles, null/negative results, case reports, data notes and more The peer review process is transparent and collaborative Your article is indexed in PubMed after passing peer review Dedicated customer support at every stage For pre-submission enquiries, contact research@f1000.com