HAP Software.

HAP analysis was implemented as a collection of software tools assembled into a computational pipeline. The components of the pipeline include: (1) a general hormone database structure for organizing hormone time-series data and experiment details including experimental protocol, participant demographics and sleep-wake state, (2) tools for applying HAP to data stored in the database, (3) tools for visualizing the hormone data contained in the database, (4) tools for visualizing the results from the HAP analysis, and (5) tools for creating summaries of individual and group data results.

The HAP analysis pipeline was written with MATLAB ver. 7.3 and contains 87 MATLAB script files. The Graphical User's Interface Development Environment (Guide) toolbox was used to create an interface for selecting, visualizing, and analyzing the hormone database structure. The HAP_Analysis program used in this analysis and the source code can be found at https://github.com/DennisDean/HapSource/releases. HAP_Analysis is a windows executable and does not require MATLAB to be installed on the computer. Text S1 contains additional program details and a summary of figures that can be generated by the program is shown in Figure S3. A video of the program executing an analysis is posted on YouTube (http://youtu.be/ggJDK7alI2M).

Cortisol Model Simulation Software.

A published mathematical model of cortisol concentration levels [33] was implemented and used to test the HAP implementation. Features of the mathematical model include randomly sampled interpulse intervals with circadian amplitude modulation. The MATLAB simulation script and a test program can be found at https://github.com/DennisDean/BrownCortisolModel. A brief description of the model and software is included in Text S1.

Experimental Data.

Fourteen healthy women were studied in an inpatient protocol [34]. Blood was sampled through an indwelling intravenous catheter every ten minutes for 24 hours and later assayed for cortisol. The first eight hours consisted of a scheduled sleep period and the remaining 16-hour wake period employed a constant routine protocol during which participants remained awake in a semi-recumbent position in dim lights and were given small regularly spaced meals [35]–[37].

Context Free Language for Identifying Hierarchically Organized Pulses (Step 1).

The CFL was motivated by the observation that a rise or fall in the cortisol concentration profile can be due to an embedded set of pulses. This multi-scale embedded feature can be represented as a hierarchical structure. The detailed construction and steps involved in a CFL example of embedded parentheses is included in Section A of Text S1. The details of the CFL for HAP are presented here.

Let be the set of the CFL that can identify hierarchical rises and falls in a time-series representing cortisol concentration measurements. is composed of a grammar that can generate strings to represent any valid cortisol concentration or other pulsatile time-series; these strings can be transformed into a computation for recognizing the CFL directly from the pulsatile time-series.

The CFL grammar that generates these strings in the language is composed of a quadruple, where a quadruple is a set of four definitions, that includes the alphabet , the set of terminals , the set of rules , and the start symbol . The components of the CFL grammar are shown in Equations (1)–(4).(1)(2)(3)(4)

This CFL provides a symbolic and compact representation of hierarchically organized pulses. Thus, the CFL is a way to represent pulsatile time-series by discrete sections or tokens, where each section describes features present in the data as defined in the language's alphabet.

Nadir Selection Algorithm Specification (Step 2).

CFL and Nadir Selection Algorithm Relationship. The CFL described above is a compact representation of pulsatile data and can be used with time-series. However, direct application is problematic since it is not possible to deterministically identify the next valid CFL production when examining a portion of the time-series from left to right (i.e., earlier to later in time). Additional information is required to determine the next valid production according to the theory of computation [26], [28]. The nondeterministic nature or use of a push-down automata (i.e., embedded parentheses example in Section A of Text S1), which is the usual means of analyzing a CFL, may require more computational resources than a deterministic algorithm, and therefore computation speed can be slow. Instead the inherent structure in matched parentheses can be used to construct an efficient and deterministic algorithm for recognizing the CFL. For these pulsatile data, a deterministic partitioning algorithm is used. The idea behind our specific partitioning algorithm, the Nadir Selection Algorithm, is that processing the entire time-series simultaneously allows matching pairs of rises, peaks, and falls in the data to be determined recursively. Thus, the Nadir Selection Algorithm is a time-series parser for converting the data into a symbolic representation. In addition, since we observed that the relationship between time and amplitude of nadirs in 24 hour cortisol profiles varied substantially by individuals (details below), we conjectured that this context free representation of pulsatility could be useful as a non-parametric approach to represent the multi-scale relationship of features (e.g., nadirs and peaks) in a pulsatile time-series. Note that an equally valid algorithm anchored on peaks could have been derived. The nadirs in the time-series were selected to anchor the partitioning algorithm since hormone pulse start times are traditionally associated with nadirs.

Nadir Selection Algorithm. The Nadir Selection Algorithm is based on a recursive algorithm for identifying and classifying peaks and troughs within the time-series data. The first step of the algorithm is to identify local peaks and nadirs. The algorithm is recursively called with nadirs as input until no rises or falls in the data are identified (Figure 3). The recursion results are used to reconstruct the productions (i.e., rises and falls) required to generate the time-series. Quantitative comparisons identify relative changes within the time-series within this hierarchy of recursion results. The quantitative comparisons include expressions for identifying concentration rises and falls, plateaus, flat regions, and missing data. The recursive algorithm was designed to exploit parallelism in the data and results in an efficient method for processing data [27], [38]–[40].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Qualitative time-series representations for a single participant (ID6).

The dotted vertical lines with lower case Roman numerals represent the same time points in the Time-Series, the Pulsicon, and the Production Graph for participant ID6. Shaded rectangles labeled with uppercase roman numerals represent the same temporal interval in the Time-Series, the Pulsicon, and the Production Graph. Time-Series: The 24-hour cortisol time-series. Nadirs identified in each of the four HAP iterations are connected with lines. The terminal HAP nadirs are identified with black circles. Pulsicon: The pulsicon for the time-series. Production Graph: Each rectangle represents productions: the application of rules defined in the CFL The five rectangular boxes represent the key features of the time-series. The first two rectangles (Labeled 1 and 2) show that there is a decreasing portion prior to the main portion of the cortisol time-series with the “>” in rectangle 2 (e.g., before left black circle). The main portion of the cortisol time-series is represented by the center rectangle (Labeled 3) with the production symbol (S). The production symbol (S) is a variable that can be replaced with any of the rules defined in the context free language. This section contains a hierarchy of embedded pulses. The last two rectangles (Labeled 4 and 5) from the production graph show that the cortisol time-series is rising after the main portion (rectangle labeled 3) of the cortisol time-series (e.g., after right black circle).

https://doi.org/10.1371/journal.pone.0104087.g003

Let , where the time-series is composed of the time vector and the concentration vector . The time vector is the set of concentration sampling times , assumed to be sampled at a constant interval , and may have missing data points. Time-based equations are presented in equations (5) and (6).(5)(6)Let be the set of hormone concentrations values that correspond to the sampling times as shown in equation (7).(7)

A comparator operator which is a function of the concentration data and a set of comparators is defined in equations (8) and (9). The comparator is a tuple where each entry is a member of the set of simple comparators defined as . Specific pairs of operators are selected that correspond to visually relevant features of hormone concentration data relative to the center point, namely nadirs , peaks , rising plateaus , descending plateaus , and flat regions . Details regarding the interpretation of the comparator operators are described in Section B of Text S1 and the comparator operators are summarized in Figure S4.(8)(9)

Let correspond to the features that are identifiable in the data as shown in equation (10):(10)where each feature corresponds to the following one of the 9 operators described above.

Let correspond to the features identified in the data by applying the comparator operator with the set of operators to the concentration data .

Let and correspond respectively to the times and concentration values of the q features identified in . The features that identified corresponding times and corresponding concentration values are defined by equations (11)–(13).(11)(12)(13)

A family of recursions is defined by recursing on a subset of the identified features. A subscripted S is used to designate the selected features with corresponding times and corresponding concentration values as shown in equations (14)–(16).(14)(15)(16)

The subset is defined to be the nadirs for the purpose of partitioning the hormone time-series into distinct pulsatile regions as described in equation (17).(17)

Similarly, a partitioning from peak to peak can be defined by selecting the peak as the subset.

Algorithm Interpretation. The identified nadirs have both a mathematical and physiological interpretation. At each level of iteration, two adjacent vertical lines in Figure 3 identify a nadir-nadir interval with a single intervening peak; these HAP-selected nadirs identify embedded rising and falling concentration segments. Mathematically, the same concept is referred to as the forced state of the system [41]. Nadirs selected at multiple iterations are identified as fixed points in the sequential dynamical systems literature [42]–[44]. The selected nadirs correspond to the start of an increase in cortisol concentration and are likely to correspond to the approximate timing of a signal to secrete cortisol.

Computing Accumulation Rates, Dissipation Rates, and Inter-nadir Intervals. The extracted hierarchical features provide a concise definition of pulsatility that can be used to calculate traditional hormone pharmacokinetic measures. In this section, we present definitions of accumulation rates, dissipation rates and inter-nadir intervals derived directly from features extracted from HAP. These metrics are analogous to secretory rates, clearance rates, and interpulse intervals, respectively,

The change in cortisol concentration rate over a finite time interval is equal to the starting concentration level plus the cortisol secreted during the interval minus the cortisol eliminated during the interval. The cortisol generated during the interval is equal to the secretory rate during secretion multiplied by the change in time . Similarly, the cortisol dissipated is equal to the clearance rate multiplied by the time cortisol is being cleared from the change in time . This relationship is shown in equation (18).(18)

For a simple pulse, the duration of secretion is equivalent to the rise duration while duration of clearance is equivalent to the fall duration, as in equation (19).(19) and are the end and start of the rise interval, respectively. Similarly, and are the end and start of the fall interval, respectively. For simplicity, secretion and clearance are assumed to be temporally distinct.

Consider a subset of features from selected at iteration where every entry is a nadir (N), decreasing to flat (DF), or flat to rising (FR). Then there exists a single peak (P), rising to flat (RF), or flat to decreasing (FD) within that provides sufficient information to estimate the first order secretory and clearance rates. For simplicity, consider a subset of features that contains three entries sequentially consisting of a nadir, peak, and nadir as shown in equation (20).(20)

If we assume that the effective secretory period corresponds to the rising portion and the clearance period corresponds to the descending portion of the interval, then equations (11)–(13), (18), and (19) can be used to compute the first order secretory rate , infusion rate , and inter-nadir interval directly from HAP outputs as shown in equations (21)–(23).(21)(22)(23)

For the simple example above, there is only one secretory rate, one clearance rate, and one interpulse interval because there are no embedded pulses. For a subset of features where each entry , then there will be secretory rates, clearance rates, and interpulse intervals where is equal to the number of local nadirs minus one. We define the hierarchical secretion rates as accumulation rates, the hierarchical clearance rates as dissipation rates and hierarchical interpulse intervals as inter-nadir intervals.

Qualitative Features Present in the Production Graph (Step 3).

The CFL production graph is presented as a way to describe the qualitative features present in a cortisol time-series. The CFL production graph is a summary of the steps in the language required to represent the data. The goal of the graph is to better visualize qualitative features such as the relations of embedded pulses and could be used to integrate graph-based analysis algorithms in future analytical work.

The production graph for the example in Figure 3 is organized in a sequence of 5 rectangles that form a high level description of the time-series from left to right. Three rectangles - labeled 1, 3, and 5 in the production graph - contain the productions required to generate their corresponding pulsicon sections. These three rectangles are expanded to show the productions required to produce the pulsicon section that represents this cortisol time-series. Two rectangles (labeled 2 and 4) identify whether the data are rising or falling relative to the main section (labeled 3). Both points in time and temporal intervals are identified to reinforce the relationship between the time-series, the pulsicon, and the production graph.

Text Representation of a Single Time-Series (Step 4).

HAP analysis generates a pulsicon or text description of the time-series. A pulsicon is a complete record of rises and falls within the time-series and can be used to evaluate the relationship among embedded pulses. Pulsicons are composed of a set of characters . A “(” indicates the start of rise and a “)” indicates the end of a decreasing pulse. The peak between two nadirs is identified with a “:”. The “>” symbol indicates a decreasing section of the time-series and the “<” symbol indicates an increasing section of the time-series. Both “>” and “<” describe data points at either the start or the end of the time-series. A pulsicon is generated for each step in the algorithm and at each hierarchical iteration. The pulsicon generated from the production graph described above is shown in Figure 3.

The pulsicon text representation of a single pulsatile time-series is composed of the same alphabet used in the CFL and is generated from the production graphs using standard production techniques for generating text strings from a graph representation [26]–[28]. The pulsicon provides a structure for summarizing the quantitative attributes of the time-series in a way that can be used by automated processing routines. For example, the pulsicon provides a common interface for automatically generating summary figures about the specific data set (e.g., Figures 4–5). Section D of Text S1 contains (i) an additional pulsicon example and (ii) an example that demonstrates the relationship between production graphs and pulsicons (Figure S5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Accumulation and dissipation rates for a single participant (ID6).

(A) Amplitude as a function of the rise or fall duration extracted from the first iteration of the HAP algorithm for participant ID6. Concentration rises (changes between local nadirs and subsequent local peaks) are shown with blue circles. Concentration falls (changes between local peaks and subsequent local nadirs) are shown with green triangles. Regression lines are shown separately for the rises (blue circles) and falls (green triangles). (B) A histogram of the instantaneous accumulation rates ( = rise amplitude/rise duration) from Panel A. (C) A histogram of the instantaneous dissipation rates ( = fall amplitude/fall duration) from Panel A.

https://doi.org/10.1371/journal.pone.0104087.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Multi-scale analysis of features identified with HAP from a single participant (ID6).

Each analysis iteration for participant ID6 is shown with a different symbol: 1-circle, 2-square, 3-up triangle, 4-down triangle. The symbol color indicates the timing of the event with the color code in right-most vertical panel: the sleep episode (blue), the first 8 hours of wake (green), or second 8 hours of wake (white). The dashed lines connect median value between iterations. (A) Cortisol accumulation rates for nadirs identified at each iteration. Accumulation rates for the first iteration are analogous to secretory rates. (B) Cortisol dissipation rates for nadirs identified at each iteration. Dissipation rates for the first iteration are analogous to clearance rates. (C) The nadir intervals identified at each iteration are shown. The event intervals at the first iteration are analogous to the inter-pulse interval. (D) The relationship between an accumulation rate and the immediately following dissipation rate.

https://doi.org/10.1371/journal.pone.0104087.g005

HAP_Analysis Verification.

HAP analysis converged for each simulated data set. The program identified rise time and inter-pulse intervals for 97% percent of the pulses (116 of 120, Table 1, Figure 7). The pulses missed by the algorithm had secretion amplitudes of less than 0.27 ug/dL. The amplitude of the missed pulses was not sufficient to impact the simulated time series profile. We would not expect HAP to identify these low amplitude events.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 7. HAP analysis of simulated cortisol test set.

(A–F) Simulated data generated from the data sets shown in Fig. 4 panels A–F. Simulated data are identified with blue markers. Peaks identified during the nadir selection algorithm are in red. Nadirs identified at a given recursive step are joined with a line, where line color represents a different recursive step.

https://doi.org/10.1371/journal.pone.0104087.g007

Accumulation and Dissipation Rates.

The accumulation rate and dissipation rates are calculated from the first HAP iteration (Figure 4) for participant ID6. The median accumulation and dissipation rates across all points are 0.067 and 0.059 µg/(dL min), respectively for this individual. Fitting a regression line to the rise points results in a 0.322 µg/(dL min) accumulation rate (R² = 0.78). Fitting a regression line to the falling points results in a 0.098 µg/(dL min) dissipation rate (R² = 0.71). . Fitting regression lines allows computing average accumulation and dissipation rates for the time-series. The slope of the regression is analogous to secretion and clearance rates.

Multi-Scale Analysis.

Using equations 17–19 the accumulation rate, dissipation rate, and inter-nadir interval at each iteration for participant ID6 are shown in Figure 5 A–C. At each iteration, the accumulation and dissipation rates and the point spread decrease. This decrease in point deviation is due to the change in time-scale at each iteration. The fastest accumulation rates occur at the end of scheduled sleep time and many of the fastest dissipation rates occur during sleep. The dissipation-accumulation plot (Figure 5D) shows an approximately log-log relationship between accumulation and dissipation across alliterations. The multi-scale analysis provides a way to assess the effect of a sequence of pulses on concentration profiles.

Accumulation Rates and Dissipation Rates.

The average accumulation rate for the 14 participants was 0.20 µg/(dL min) and ranged between 0.07 and 0.33 µg/(dL min) when all the points from the first iteration are included (Table 3). The R² value for each participant, as computed during the linear regression of rises and falls in the data, ranged between 0.14 and 0.86. The average dissipation rate was 0.08 µg/(dL min) and ranged between 0.05 and 0.10 µg/(dL min); its R² value ranged between 0.28 and 0.73. There was a positive correlation between the accumulation and dissipation rate (Pearson = 0.44).

Multi-Scale Analysis.

HAP analyses produces results on at least two time scales: the pulse scale in the order of minutes and the ultradian scale (less than 24 hours). For the 9 participants that converged in 3 iterations, the pulse scale and ultradian scale components are, 0.67 and 2.67 hours, respectively. In the 5 participants that converged in 4 iterations, HAP computed a pulse scale and two ultradian scale components corresponding to 0.67, 2.33, and 8.17 hours. We postulate that the participants whose hormone time series converge in 3 vs. 4 iterations may have differences in cortisol control. These inter-participant differences appear at the second iteration index in the pulsicons (Table 2); pulse clustering (i.e., start of a new pulse before the effects of previous pulses are completely dissipated) occurs in those who converge in 4 iterations. Further details of how pulsicons represent these differences are included in Text S1.

Language-Based Pulse Analysis Technique and Analysis Pipeline.

The ability to automatically describe pulsatility formally is a novel HAP analysis feature. This feature facilitates the development of an analysis pipeline, and it is intrinsically tied to visualizations. This theoretical framework for HAP is implemented as an extendable analysis pipeline whose design is motivated by automated methods derived from CFLs and current hormone analysis programs [15], [16] and with the aim of translating HAP methods into use. Our pipeline provides support for analyzing data collected in a database, which differs from many current interfaces that analyze data from one participant at a time [15], [16].

Hierarchically Organized Results.

The hierarchically organized results from HAP analysis motivated a system of visualizations for summarizing pulsatility present in a time-series. For example, plots that illustrate the multi-scale nature of pulsatility (e.g., Figure 5) both provide a way to evaluate how features and parameters of the system are changing at different time-scales and provide insight into the number and organization of pulses. Visualizations for identifying the embedded nature of cortisol pulsatility (e.g., Figure 3) are another novel application of CFL production graphs. Pulsicons provide a new tool for understanding the pulsatility embedded within a hormone time-series through the ability to view the time-series as a text string. The pulsicons provide a framework to characterize and to potentially search time-series; this is a topic for future work. Linking HAP analysis output to visualizations is designed to facilitate verifying, validating, and analyzing results and to empower the user to explore features present in their data. The interface between generating HAP results and generating graphs is designed to employ grammar generated visualizations [50], which will allow for new visualizations to be generated interactively by the user.

Expression of an Eight Hour Ultradian Rhythm in a Subset of Participants.

HAP identified an 8-hour periodicity in the subset of participants that converged in 4 iterations and that was not present in the group that converged in 3 iterations. The approximate 0.7 and 2.5 hour signals were identified in all participants by HAP; Veldhuis identified similar periodicities in 24-hour cortisol concentration profiles [13]. However, Veldhuis did not report an 8 hour periodicity (See Multi-scale Analysis section in Group Results). We conjecture that the inter-individual difference in ultradian signal could be due to individual differences in how time-of-day affects the Hypothalamic-Pituitary-Adrenal (HPA) axis or in the effect of the 8-hour sleep episode on cortisol pulsatility. Our conjecture that sleep is implicated is supported by a recent article that reports ultradian components in rodent sleep [53].Future work including more than 24 hours of data and more than one sleep episode per person will be needed to define the characteristics of this finding.

Qualitative Structure of Cortisol.

In this paper, we report an objectively derived qualitative structure of cortisol pulsatility that is applicable to a range of cortisol profiles and is represented visually as a production graph and text string. These qualitative structures are representation patterns of pulsatility. Since patterns of hormone pulsatility affect responsiveness to the hormone [54]–[56], it is important to understand and define the underlying patterns. We hypothesized that the qualitative structure of cortisol pulsatility reflects the effect of sleep on cortisol production and is associated with ultradian signals.