When we define Machine Learning, we describe it as is a computer science field in which an algorithm learns from an environment or situation in an automatic way. The obtained result is frequently called “the model”, capable to respond to a given stimulus. In order to organize the learning process, one of the most used techniques is supervised machine learning. In this case, the technique learns from labelled data correctly representing the environment. In other words, the system learns the features or important descriptive characteristics of an input (e.g. imagine an orange, features from the input orange can be round shape, orange colour,…), which is labelled (the input is labelled as “orange”) so that the system learns in the future features and could determine automatically the solution of the problem (identification of “oranges” among data) .
Machine learning (ML), as part of the Artificial Intelligence (AI) domain, is present in our lives for several years. When ML comes into our minds, we think that it is focused on understanding and helping very different contexts: business tendencies (stock market), industrial processes (object recognition), and recommendations from web services (online services such as clothing, music…).
The use of ML techniques in the context of medicine and health is not new. The scenario and applicability in the medical field are very important but we find some barriers. Health care Professionals (HCP) believe that at some point ML will replace them but from the engineering point of view, this is a real fallacy. There is an important challenge today, which is creating the awareness that ML will help HCP to make better decisions. ML will not replace them, on the contrary, ML will improve the quality of diagnostic and medical service. This will make a direct impact on the patient, who will obtain a better quality of life by having a correct therapy prescription. With that said, what is Machine Learning?
In this short article, we explain the case of Parkinson’s Disease (PD), but it can be extrapolated to many cases in the medical field. The case of Parkinson’s is a clear example of why it is useful and becomes necessary to use ML in clinical practice.
In a classical visit to the doctor, the patient with PD has a very short time to report information to the neurologist. Frequently, patients have to explain what has happened in the last 6 months and what is the behaviour of the disease and all its symptoms in only 15 minutes. An additional problem is that the patient comes medicated, so they don’t show the real symptoms to the physician, who has to find out what is the severity of the symptomatology, which is very complex in PD. Moreover, the patient can suffer from cognitive impairment, difficulties speaking, anxiety, stress, or depression , making it really complex to provide a clear map of PD symptoms.
Furthermore, patients with PD are not only affected by motor problems and then the diagnostic is not dichotomic. The motor symptoms, additionally, fluctuate with different severities along the day. One can suffer from bradykinesia (slowness of movement characterized by a very slow gait), stiffness (rigidity provoked in upper and lower limbs), tremor (in hands), freezing of gait (sudden block in the gait that is correlated with falls), dyskinesia (affectation consequence of PD drugs) . Thus, it is totally impossible to know the severity, distribution, and frequency of all the symptoms in a few-minute visit.
The COVID’s pandemic scenario is seriously alarming the health services since several patients are not going to their usual visits due to the fear of becoming infected just by going out of their home, or by entering a zone, which they consider risky.
Altogether, several neurologists are claiming that medical wearable devices that monitor objectively PD symptoms are necessary to improve the evaluation process in normal life conditions (out of the hospital or physician’s office). In the last International Movement Disorders Conference (MDS2020), in the Plenary Session called “Digital Technologies for Diagnosis and Disease Monitoring” it was highlighted that there is a need to evaluate patients at home to provide objective information to the physician.
Inertial wearable systems are a must and can provide reliable and objective information to neurologists. These systems are integrating very sensible movement sensors, together with a high computation capability. The obtained data from the sensors are processed by the different embedded algorithms that could be based on ML. The question now is: how is a wearable medical device built based on ML techniques and what are the main advantages?
In the field of Parkinson Disease, the target is to learn from the movements of the patient, and the machine learning strategy consists of an algorithm that gets information, normally from a set of inertial sensors, and based on the analysis of a short period of this accelerometer signals, the algorithm determines the corresponding output. This output can be that the patient is suffering a certain symptom with a certain severity.
To do this process automatically, the supervised machine learning algorithm has been correctly trained and has learned from a labelled database that is representing the complete scenario with the different cases and movement features associated with representative symptoms. Thus, it is important to build a robust database, which covers, on one hand, all the target symptoms, for example, bradykinesia, dyskinesia, motor fluctuations, freezing of gait…this way, the algorithm will maximize the sensitivity, which indicates how well the algorithm detects a symptom. On the other hand, it is mandatory to build another inertial database with activities of daily living, such as cleaning the home, walking, being still, going by car, bus, train…then, this way the algorithm will maximize specificity, which is the indicator that says how good is my algorithm to say that this movement is not a possible symptom. Balance in this database is essential, and covering all movements is burdensome, but then, the real accuracy of the algorithm will be higher. It must be noted that many studies published in top-rated journals have only covered the movement of the patient in clinical settings, and when the algorithm is tested in real life, it does not work as expected having a very low specificity by having a lot of false positives.
To understand the labelling of a signal, let’s assume for example a triaxial accelerometer signal of 60 seconds of duration, with a sampling frequency of 100Hz, forming 3 arrays of 6000 samples, one for each axis. Then, the label is also an array from 6000 different values in correspondence to a given activity. Each value is associated for example with activities like walking, running, standing…. Figure 1 shows an example of a triaxial accelerometer and the “labelling” signal.
A possible strategy of the ML implementation is to try to learn directly the signal waveform associated with the sample by adding all the samples of the signal as input parameters. This, however, might involve a high degree of complexity, as not all samples are relevant and this method only takes more computational burden to the algorithm, thus, it is important to apply an alternative strategy. From the huge and unfiltered database, it is aimed to extract key information from the signal (the features associated with the different movements and symptoms). To do so, the signal is split into the so-called windows, short periods of the signal, and from which we extract its features to which we will associate the given label (see Figure 2).
From each window, we can extract features that can be time-based characteristics, such as the amplitude of the signal, maximum, minimum… or frequency-based characteristics, for example making a Fourier Transform from this part of the signal, extracting frequency information, which is relevant to detect the gait or other repeated human movements.
In case the number of features is extremely high, then we should select the most relevant by the application of well-known reduction methods. For example, if we have the standard deviation and the variance, probably one of the two features could be removed as the information of one is included in the other. In case we have much more features, we can use algorithms to select features, such as r-Relief . Once we have an optimal set with all the selected significant features from a window of signal, then this window is labelled with the target to be determined by the machine learning algorithm.
A good example is the detection of the Freezing of Gait (FoG) Parkinson’s Disease symptom [4,5]. In this algorithm, it is selected a series of features to detect if a window is a FoG or not. Then all the windows, from which were extracted several features, were labelled as FoG= ‘1’ if this window contained a signal from a FoG episode or FoG= ‘-1’ if the current window did not contain info from FoG episodes. All these windows are then trained with different methods. From each method, a series of hyperparameters must be configured, such as the number of layers of a neural network, or the number of trees in the Random Forest method.
Then, the obtention of a model is based on long iterative mathematical processes where certain initial conditions are set and based on the selected hyperparameters, the model is built and tested over a blind dataset to get performance results that will be analyzed later. Several programs ease this long and complex process such as Weka, Matlab, R, or code done with Python. After analyzing the performance of an algorithm by comparing outcomes such as sensitivity, specificity, positive and negative predictive value, f-score, accuracy…then “the model” is obtained.
To sum up, it is crucial to use machine learning techniques to generalize the algorithms as threshold-based algorithms do not work well in complex detections. However, there is a bad tendency in low-cost devices to use a simple threshold to detect something, leading to a high rate of false positives or false negatives. In the case of Parkinson’s Disease, there are many examples using support vector machine/regression [4,6] or even deep learning [7,8] that although the algorithms could be complex, the generalization of the problem is promising and it becomes a reality to test it in real life. Once the model algorithm is extracted, one can embed the system in real-time as it is done with STAT-ON, a medical device that monitors motor symptoms of PD , or one can send the information to a server [10–12]. Accordingly, the rise of inertial wearable medical devices is enabling health professionals to know what is happening in real life.
The use of medical devices to evaluate a patient with Parkinson’s Disease signifies the start of a new paradigm. Although this sounds challenging, the truth is that many neurologists are considering medical devices as a very helpful tool in clinical practice , which contributes to a better treatment prescription and, ultimately, to an improvement in the quality of life in these patients.
A clear example is the use of advanced therapies in Parkinson’s Disease. It has been shown that advanced therapies clearly improve the quality of life in patients [14–16]. However, there is a noteworthy number of infra-diagnosed patients with advanced symptoms that could be potential candidates to use these device-assisted therapies . The lack of information of health professionals could result in patients not having the correct therapy, and this way, medical devices based on machine learning could play a very important role.
Machine learning can be complex to understand, but, the truth
, is that these automatic methods to understand large amounts of data are changing the world, not only in health, but in economics, security, biology, or business. It is also very interesting to see how many machine learning devices are being introduced in the USA in the latter years . The appearance and application of machine learning in health have arisen significantly the new “de Novo” devices introduced in the USA.
The use of machine learning will help professionals in any field to have more information about the context and will enable them to make better decisions. The barriers of skepticism, however, must be broken and, still, a lot of work has to be done.
The author of this article would like to thank Professor Joan Cabestany for his support in the review of this article.
1. Bishop C. Pattern Recognition and Machine Learning. Jordan M, Kleinberg J, B. Scholkopf, editors. New York: Springer; 2006.
2. Jankovic J. Parkinson’s disease: clinical features and diagnosis. J Neurol Neurosurg Psychiatry. 2008;79: 368–76. doi:10.1136/jnnp.2007.131045
3. Robnik-Sikonja M, Kononenko I. Theoretical and Empirical Analysis of ReliefF and RReliefF. Machine. 2003;53: 23–69.
4. Rodríguez-Martín D, Samà A, Pérez-López C, Català A, Moreno Arostegui JM, Cabestany J, et al. Home detection of freezing of gait using support vector machines through a single waist-worn triaxial accelerometer. PLoS One. 2017;12. doi:10.1371/journal.pone.0171764
5. Samà A, Rodríguez-Martín D, Pérez-López C, Català A, Alcaine S, Mestre B, et al. Determining the optimal features in freezing of gait detection through a single waist accelerometer in home environments. Pattern Recognit Lett. 2017; doi:10.1016/j.patrec.2017.05.009
6. Samà A, Pérez-López C, Rodríguez-Martín D, Català A, Moreno-Aróstegui JM, Cabestany J, et al. Estimating bradykinesia severity in Parkinson’s disease by analysing gait through a waist-worn sensor. Comput Biol Med. 2017;84: 114–123. doi:10.1016/j.compbiomed.2017.03.020
7. Camps J, Samà A, Martín M, Rodríguez-Martín D, Pérez-López C, Alcaine S, et al. Deep Learning for Detecting Freezing of Gait Episodes in Parkinson’s Disease Based on Accelerometers. In: Rojas I, Joya G, Catala A, editors. Advances in Computational Intelligence. Cham: Springer International Publishing; 2017. pp. 344–355. doi:10.1007/978-3-319-59147-6_30
8. Sigcha L, Costa N, Pavón I, Costa S, Arezes P, López JM, et al. Deep Learning Approaches for Detecting Freezing of Gait in Parkinson’s Disease Patients through On-Body Acceleration Sensors. Sensors. 2020;20: 1895. doi:10.3390/s20071895
9. Rodríguez-Martín D, Pérez-López C, Samà A, Catala A, Cabestany J, Rodríguez-Molinero A. STAT-ON: a Wearable inertial system to objectively evaluate motor symptoms in Parkinson’s Disease. Movement Disorders Society Conference. 2019.
10. Joshi R, Bronstein JM, Keener A, Alcazar J, Yang DD, Joshi M, et al. PKG Movement Recording System Use Shows Promise in Routine Clinical Care of Patients With Parkinson’s Disease. Front Neurol. 2019;10. doi:10.3389/fneur.2019.01027
11. Kinesia objective assessment [Internet]. 2016 [cited 1 Sep 2016]. Available: http://glneurotech.com/kinesia/products/kinesia-360/
12. Morsy A. Patient Technology for Tracking Movement Disorders [Entrepreneurs Corner]. IEEE Pulse. 2018;9: 6–40. doi:10.1109/MPUL.2018.2858458
13. Santos-García D, López Ariztegui N, Cubo E, Vinagre Aragón A, García Ramos R, Borrué C, et al. Use in Clinical Practice of a Personalised and Long term Monitoring Device for Parkinson’s Disease: STAT ON. International Congress of Parkinson’s Disease and Movement Disorders. Philadelphia; 2020.
14. Rabie A, Verhagen Metman L, Fakhry M, Eassa A, Fouad W, Shakal A, et al. Improvement of Advanced Parkinson’s Disease Manifestations with Deep Brain Stimulation of the Subthalamic Nucleus: A Single Institution Experience. Brain Sci. 2016;6: 58. doi:10.3390/brainsci6040058
15. Drapier S, Eusebio A, Degos B, Vérin M, Durif F, Azulay JP, et al. Quality of life in Parkinson’s disease improved by apomorphine pump: the OPTIPUMP cohort study. J Neurol. 2016;263: 1111–1119. doi:10.1007/s00415-016-8106-3
16. Ciurleo R, Corallo F, Bonanno L, Lo Buono V, Di Lorenzo G, Versaci R, et al. Assessment of Duodopa® effects on quality of life of patients with advanced Parkinson’s disease and their caregivers. J Neurol. Springer Berlin Heidelberg; 2018;265: 2005–2014. doi:10.1007/s00415-018-8951-3
17. Ávila A, Pastor P, Planellas L, Gil-Villar MP, Hernández-Vara J, Fernández-Dorado A. DISCREPA study: Treatment of advanced Parkinson’s disease and use of second-line treatments in Catalonia. Rev Neurol. 2021;72: 1–8. doi:10.33588/RN.7201.2020181
18. Yiannis Mouratidis. The Rise Of Use Of Medical Devices Force FDA To Change The Rules [Internet]. 2018 [cited 17 Oct 2019]. Available: https://www.forbes.com/sites/yiannismouratidis/2018/12/09/the-rise-of-use-of-medical-devices-force-fda-to-change-the-rules/#7a9631962101