Data to diagnosis in global health: a 3P approach

In this paper, we propose novel approaches to transform data into diagnosis. As a collaborative work between our researchers and clinicians in one of the largest super-specialty hospitals in India (Amrita Institute of Medical Sciences - AIMS), we developed physician assist filters (PAFs) that are designed to transform unwieldy time series sensor data into summarized patient/disease specific trends in steps of progressive precision as demanded by the doctor for patient’s personalized condition at hand, and help in identifying and subsequently predictively alerting the onset of critical conditions. Together with the communication network and data transmission architecture, this new framework that we have designed, developed, and successfully deployed is called RASPRO (Rapid Active Summarization for effective PROgnosis). The RASPRO framework was first introduced in [1].

We begin by analyzing the existing systems that simply generate alerts every time one or more sensors cross the abnormality thresholds. Due to the humongous volume of such alerts, they are difficult to manage even in the case of hospital in-patient settings, let alone for a much larger number of remotely monitored patients. Starting from some of the initial attempts reported in [2], to more recent works such as [3‐5], and [6], the severity detection and alert generation is typically based either on predefined thresholds, or based on training of thresholds using machine learning followed by online classification of multi-sensor data. Very similar techniques of machine learning have also been used in fall detection [7, 8]. Hristoskova et al. [9] propose another system wherein patient conditions are mapped to medical conditions using ontology-driven methods, and alerts generated based on corresponding risk stratification. Even though there has been noticeable success in detection and diagnosis of specific disease conditions, most of these works have not explored the opportunity for personalized and precision diagnosis. In an extensive review of Big Data for Health, Andreu-Perez et al. [10] specifically emphasize on the opportunity for stratified patient management and personalized health diagnostics citing examples of customized blood pressure management [11]. More specifically, Bates et al. [12] discuss the utility of using analytics to predict adverse events, which could reduce the associated morbidity and mortality rates. Furthermore, Bates et al. [12] argue that patient data analytics based on early information supplied to the hospital prior to admission can result in better management of staffing and other hospital resources. One of the recent works in personalized criticality detection is reported in [13], which propose an analytical unit in which the Improved Particle Swarm Optimization (IPSO) algorithm is used to arrive at patient-specific threat ranges. To improve precision in diagnosis we also need to arrive at a balance between a completely automated system on one hand, and physician assist systems on the other. Celler et al. [14] propose a balanced approach wherein sophisticated analytics are presented to physicians who in turn identify the changes and decide on the diagnosis. This is also supported by many results including that reported in [6], wherein domain knowledge based method performed as well as other trained machine learning models. These arguments and results provide further impetus for personalized, precision, and preventive diagnostic techniques that are amenable to physician interventions.

In general, let us consider N body sensors, S₁,S₂,…,S_N with varying sensing frequencies f₁,f₂,…,f_N. The raw time series values from these sensors are converted to discrete severity level symbols by the quantizer. The number of severity levels L_i for a sensor S_i can be set based on the sensor and many other factors. We assume that different vital parameter sensors have different number of severity levels, and hence L₁, say the number of severity levels for a blood pressure sensor, could be equal to five, whereas, L₂ (say oxygen saturation levels) could be equal to seven. In our symbolic notation, the clinically accepted normal values are assigned the symbol A, while above-normal values are assigned with progressive degrees of severity as A+, A++ etc., while that of sub-normal values are assigned A-, A−− etc.; the number of “+” and “-” symbols representing degree of normal and subnormal severities respectively. Figure 2 depicts how various severity levels are arrived at in Personalization PAF severity quantizer.

The quantized severity symbols are arranged into a Patient Specific Matrix (PSM) of N rows and W columns, where N is the total number of sensors being observed, and W is a time window in which the data is summarized. The value of W can be set by a physician or automatically derived based on the risk perception of that particular patient.

The quantization breadth are decided by doctors based on the patient profile (or history), doctor’s diagnostic interest (for instance, a cardiologist may assign severity ranges differently from that of a nephrologist), severity ranges as suggested by using analytics on local hospital information system (HIS), and also based on population analytics across multiple HIS spanning multiple hospitals or even from publicly available databases such as PhysioNet [15]. Together, this approach gives ample flexibility in achieving customization in inter-patient, inter-disease, intra-patient, inter-specialty diagnosis from multi-sensor data.

The electrocardiogram is a potential indicator of cardiac events and can be exploited for personalized and precision diagnosis by varying the parametric thresholds and summarization window, based on patient profile/disease condition and associated factors. For instance taking into account the disease condition, a 3mm depression in the ST segment would be graded as A++ for an active patient having exertion related chest pain, indicating cardiac ischemia, whereas the same if occurred in a patient at rest, would be graded as A+++ with limited time of intervention (30 min), indicating cardiac muscle death. To extend the spectrum of diseases that ST segment depression would cover, a chronic hypertensive with left ventricular hypertrophy of the heart (and no chest pain) would also presumably have a continuous 3mm dip in the ST segment which does not require any interventional attention, and hence, would be graded as A/A+ (near normal) by the severity quantizer. Next, taking into account the patient profile, in sedentary workers, aged above 45 having smoking habit, with high cholesterol levels, and other associated risks the thresholds will be low (A+, A++, and A+++ would be assigned to 1–2mm, 2–3mm, and above 3mm ST depression respectively) while in highly active but risk patients with age less than 45, and no previous associated history, the levels will be high (A+, A++, and A+++ would correspondingly be assigned to 2–3mm, 3–3.5mm, and above 3.5mm respectively). Also, in the former case the summarization window W (capturing how long ST depression sustains) would be 3–4 min (more critical), whereas in the latter it would be 7–9 min.

Simple but vital parameters such as oxygen saturation levels in the body (SpO2), BP, heart rate variability (HRV), and respiratory rate variability (RRV), present in unique combinations, would facilitate differentiating between benign diseases such as interstitial lung disease/sleep apnea for which the thresholds for alert (set through the interventional time constant K_P) will be fairly high, and emergencies such as pulmonary edema/pulmonary embolism (blood clot in an artery in the lung) for which the thresholds will be kept low if any of the predisposing factors such as left heart failure, pulmonary hypertension, prolonged immobilization, pregnancy, etc. are present. Hence, the physician would preset these combinations of vitals to be looked for as sequence of symbols in the CAM. Since the number of parameters that could be picked up to indicate disease are a few, it is pertinent that stepwise precision techniques such as machine learning algorithms be used for distinguishing between closely mimicking conditions. To quote as an example is obstructive sleep apnea and chronic obstructive pulmonary disease, both of which would show similar trends in SpO2, HRV, RRV and BP. In a trial that was conducted at our hospital, we were able to achieve 99% precision in diagnosing sleep apnea from HRV using deep learning algorithm called Long short-term memory recurrent neural networks (LSTM-RNN) and is reported in one of our previous works [16]. The algorithm evaluation was done using the multi-sensor patient data from the Physionet Challenge 2000 [18], which contained annotated data from 35 patients who underwent overnight sleep study.

One of the early markers of autonomic neuropathy in epileptic patients is the discrepancy between the BP and the pulse rate of the patients. In this scenario, the severity levels of BP and pulse rate would be set accordingly (as a combination) to alert the practitioner. Suppose S1 is BP and S2 is heart rate sensor respectively. Let us say for the patient P1, μ_CAM[P1]=<A−−,A>, and for patient P2, μ_CAM[P]=<A−,A+>. In both cases, the diagnosis, alert level, and treatment vary because P1 has BP decline with no change in heart rate (critical), while P2 has a compensatory increase in heart rate, which indicates good autonomic function.

Though these are representative clinical scenarios, we found wide agreement among the doctors from other specialties too that personalization, step-wise precision, and prevention introduced through the RASPRO framework is of high utility in remote monitoring and critical alert generation.

The first step to evaluate these hypotheses is to identify datasets that are extensive, long term and critically significant. We used large time series dataset from MIMIC II [19] database, which contains multiple body sensor values from over 20,000 ICU patients. This dataset consists of ECG, ABP (Arterial Blood Pressure), Heart Rate (HR), Non-obtrusive BP (NBP), SpO2, Mean Arterial BP (MAP), and other vital signs. From this, we selected a curated set of patient and control group data that contained a long time series data followed by a critical event. We selected patients with Acute Hypotensive Episodes (AHE), which is a potentially fatal condition, found quite common in ICUs as well as caused due to postural hypotension. An AHE event is analytically identified as when MAP measurements remain below 60 mmHg for more than 30 min. This is a potentially fatal event and requires immediate intervention. We also made sure that the dataset provides uninterrupted MAP signal with a minimum sampling rate of 1 per minute, over at least 3 h for both the event-patients as well as the control group. We selected a group of 35 patients (called group H) who had AHE during some time during their stay in ICU, and another 35 patients (called group G) who did not have AHE during their ICU stay. This dataset was selected from the PhysioNet [15] challenge 2009 [20]. The H dataset also had a time marker t₀, after which AHE occurred in that patient within a one-hour window. Since the data was obtained from publicly available sources, we did not require getting prior approval of IRB for this work.

The first task is to measure the replaceability of the original time series data with the quantized symbols and consensus motifs. To evaluate this, the H and G group time series data comprising of Mean Arterial Pressure (MAP), of length 60 min after t₀ are modeled as feature vectors of length 60. These vectors are called original time series (OTS), and are used for training an SVM model for classifying the data as having AHE or not. The vectors belonging to AHE were labelled as H, and G otherwise. After using OTS, we then generate quantized time series (QTS) vectors with different quantization breadth. The quantization breadth (denoted by B) are varied as 5, 10, 15, and 20. For instance, when B=10, each of the OTS MAP values between 60 mmHg and 50 mmHg are quantized into the same severity symbol, say “A-”, whereas for B=5, the symbol “A-” quantizes all OTS MAP values between 60 mmHg and 55 mmHg. These vectors are used in similar manner to first train and then test the SVM model. Finally, we generate the corresponding motif time series (MTS) for each of the QTS, with varying the summarization time window W as 5, 10, and 15. The value of W corresponds to the time window in which all the severity symbols in the QTS are converted to a single consensus symbol. A comparison of OTS, QTS, and MTS is done using the statistical measure of binary classification, the F-score. An F-score (also called F₁-score) is calculated as:

The next evaluation parameter of RASPRO is to identify if a priori motif series could predict future disease condition, and thereby aid in preventive intervention. For this, the H and G group time series data comprising of Mean Arterial Pressure (MAP), of length T minutes prior to t₀ is modeled as a T-long feature vector (OTS). These vectors are used for training (using 70% data, with 5 fold cross validation) and testing (using 30% data) an SVM model for predicting them as AHE or not, where patients belonging to H group are annotated as having AHE and G group patients are annotated otherwise. In effect, we try to classify sensor data prior to AHE event as a predictor for ensuing AHE condition. Since G group data did not have a time marker t₀ we selected a random but continuous time series of length T from each of the G group patients. SVM was selected due to its widely accepted performance in classification problems involving multiple features, although we might obtain comparable results using other classification techniques too.

The backward offset time T (from t₀) is varied as 30, 60, 90, 120, 150, and 180 min as an expanding window. In the next step, the raw feature vectors are quantized using severity quantizer to form a quantized time series (QTS). Once again, the quantization breadth B is varied as 5, 10, 15, and 20. In the third step, the QTS are summarized and motifs extracted to form RASPRO Motif Time Series (MTS), with varying observation time window sizes W: 5, 10 and 15 min. The QTS and MTS are then given as input to train and test the SVM model (one for QTS and another for MTS) for predicting AHE before it’s onset.

The third hypothesis aims to find out if there are patient specific custom severity levels, and summarization frequencies, which if optimized could lead to better accuracies in diagnosis. For this, we further analyze our earlier results. We observe from Figs. 7 and 8 that by selecting different severity quantization breadth (B) and through varying the summarization window size (W), we are able to predict the onset of AHE with higher F1-score. This supports an argument for using disease and time-specific B and W values for achieving better accuracy in classification problems. We observe very similar results in Fig. 4, which shows that by choosing optimized W and B values, the machine learning models can perform better in classification problems too. These results further support our third hypothesis, that there exists an opportunity for personalization at least at disease specific and time specific level. Though the above experiments using AHE are only representative of how step-wise precision, personalization, and prevention can be achieved using RASPRO, the practitioners as a whole agree that in wide-ranging scenarios patient-sensor-disease-time specific severity levels need to be defined that is both practical to manage alerts as well as effective in identifying emergencies.

These medical benefits of RASPRO framework would contribute directly to fulfill the primary goals of remote health monitoring in global health scenario. We call these as the 3A benefits in short, which stands for: availability, accessibility, and affordability.

These advantages would help bring quality healthcare to millions of people who are currently under-served in the global health scenario. We are readying for large-scale deployment of the RASPRO framework including the 3P RASPRO-PAF analytical tools using a network of more than 45+ telemedicine nodes (as shown in Fig. 10) and remote health centers across the Indian sub-continent, which are connected to the AIMS hospital.