Modeling and Analysis

Objective activity parameters track patient-specific physical recovery trajectories after surgery and link with individual preoperative immune states

Fallahzadeh, Verdonk et al.

Study Materials Modeling and Analysis

Supplementary Figure 1. Consort Chart Supplementary Figure 2. Gating strategy

Supplementary Figure 3. Prediction of time since surgery

Supplementary Figure 4. Performance comparison of the clock of recovery vs the univariate models Supplementary Figure 5. Correlation between preoperative immune features and postoperative physical recovery

Supplementary Figure 6. Prediction of recovery surrogates from preoperative activity and sleep metrics Supplementary Table 1. Accelerometry features

Supplementary Table 2. Mass cytometry panel

Supplementary Table 3. Coefficients of mass cytometry Canonical Correlation Analysis model References

Study Materials

Patient exclusion criteria

The following conditions rendered patients ineligible for study participation: age < 18 years, infectious disease within the last month, immune-suppressant therapy within the last 2 months, chronic medication with immune-modulatory effects, major surgery within the last 3 months or minor surgery within the last month, history of substance abuse, pregnancy, autoimmune disease, organ dysfunction resulting in significant clinical impairment, and active malignancy.

Patient-reported outcome tools

Patient-reported clinical outcomes were captured with the Surgical Recovery Scale (SRS) and an adapted version of the Western Ontario and McMaster Universities Arthritis Index (WOMAC) [1]. The SRS has specially been developed and validated to quantify fatigue and resulting impairment of daily activities in surgical patients [2]–[5]. The SRS is sensitive to change and correlates with course and severity of postoperative complications [1], [6]. Patients were asked to use Likert scales to answer thirteen questions assessing fatigue and resulting impairment of performing different daily activities. A compound score is derived ranging from 17 to 100 (worst to no fatigue/impairment). An adapted version of the WOMAC scale was used because not all questions are applicable in the context of surgery. The scale is well validated and reliably tracks functional impairment over time [1], [7]–[11]. Patients were asked to use an 11-point numerical scale to rate pain at night, at rest, while weight-bearing, and while walking on a flat surface. A compound score is derived ranging from 0 to 40 (no to worst pain). Patients were also asked to rate functional impairment of the operated leg while lying, sitting, raising from the bed, raising from the chair, standing, and walking on a flat surface. A compound score is derived ranging from 0 to 60 (no to worst impairment).

Accelerometry

Participants were instructed to continuously wear a clinical research grade smartwatch (ActiGraph GT9X Link) for a total duration of 45 days (5 days prior to surgery and the following 40 days, not including the day of surgery). The three-dimensional wrist acceleration data was recorded at 30 Hz sensing frequency and stored locally on the device. The de-identified HIPAA compliant patient data was transmitted into a secure cloud server after the completion of the trial. ActiGraph's data analysis software platform ActiLife v6.13.3 was used for feature extraction. Default values were used for all settings and parameters except for Epoch Length (60), and the selection of the Tudor-Locke algorithm sleep detection, and the Cole-Kripke Sleep/Wake algorithms.

Mass Cytometry Analysis

Whole-blood samples were collected in sodium-heparinized tubes 1hr prior to surgery. Within 30 minutes of phlebotomy, samples (1 mL) were stimulated with GMCSF (100 ng/mL, R&D Systems, Minneapolis, MN), IFN- (100 ng/mL, PBL Assay Science, Piscataway, NJ), a mixture of IL-2, -4, and -6 (each 100 ng/mL, R&D α

Systems), LPS (1 ug/mL, InvivoGen, San Diego, CA), and PI (Cell Stimulation Cocktail, Invitrogen, San Diego, CA) or left unstimulated, were fixed in Smart Tubes (Smart Tube Inc., San Carlos, CA), and then immediately stored at −80 °C.

The tissue trauma associated with surgery triggers the release of inflammatory mediators, such as alarmins and inflammatory cytokines, which in turn results in the systemic alteration of innate and adaptive immune cell signaling responses. Our group and others previously showed that activation of the MyD88 pathway (including the P38 and NFkB signaling) in myeloid cell subsets, and the JAK/STAT (notably STAT1, STAT3 and STAT5) signaling and ERK/MAPK signaling in myeloid and lymphoid (e.g., CD4+ and CD8+T cell subsets) cell subsets are core elements of the human endogenous immune cell signaling response to surgery [1]. The rationale for selecting our ex-vivo stimulation panel is to allow assessing the integrity of canonical innate and adaptive immune signaling responses that are endogenously altered in response to surgery. As described in detail in [12], chosen stimulations elicit such cell-type specific responses in a robust and reproducible fashion.

Specifically, stimulation with LPS activates MyD88 signaling responses primarily in myeloid cells; the combination of IL-2/IL-6/IFN-a activates JAK/STAT (predominantly STAT1/STAT3/STAT5) signaling responses in both innate and adaptive immune cells, while GM-CSF triggers STAT5 signaling responses in myeloid cells selectively. PMA/Ionomycin was also selected as it allows robust activation of calcium- dependent signaling (notably ERK/MAPK signal) responses, notably in adaptive immune cell subsets.

After thawing and erythrocyte lysis, samples were barcoded and stained with surface and intracellular antibodies using standardized protocols [13]. In brief, the mass cytometry antibody panel included 28 antibodies that were used for phenotyping of immune cell subsets and 11 antibodies for the characterization of intracellular signaling (Supplementary Table 2). Antibodies were either obtained preconjugated (Fluidigm, Inc.) or were purchased as purified, carrierfree (no BSA, gelatin) versions, which were then conjugated with trivalent metal isotopes utilizing the MaxPAR antibody conjugation kit (Fluidigm, Inc.).

Whole blood samples were subjected to erythrocyte lysis using Thaw-Lyse Buffer (Smart Tube, Inc., San Carlos, CA) and isolated leukocytes from each sample were treated (barcoded) with a unique combination of 3 palladium isotopes (Trace Sciences, International, Wilmington, DE) in 0.02% saponin (Millipore-Sigma, St.

Louis, MO). After barcoding, cells were pooled and treated in aggregate with 1:100 Human Fc block (Biolegend, San Diego, CA), stained with surface antibodies, then permeabilized with methanol and stained with intracellular antibodies and finally treated with an iridium-based DNA intercalator (Fluidigm, Inc., South San Francisco, CA). All antibodies used in the analysis were titrated and validated on samples that were processed identically to the samples used in the study.

Barcoded and antibody-stained cells were analyzed on the mass cytometer. In order to minimize experimental variability, samples were barcoded, stained, and run simultaneously on the mass cytometry instrument.

Barcoded samples were analyzed at a flow rate of 600–800 cells/s. The output FCS files were normalized∼ and de-barcoded using MatLab-based software. The resulting FCS files were uploaded to the Cell Engine (https://cellengine.com, Primity Bio, Fremont, CA) flow cytometry analysis platform. Manual gating was performed according to the gating strategy in Supplementary Figure 2. The following cell types were included in the analysis: neutrophils, CD27+ Bmem cells, CD27−Bnaive cells, CD38+ CD24+ plasma and CD38+ CD27+

transitional B cells, CD56hiCD16− NK cells, CD56loCD16+ NK cells, CD4+CD45RA−T cells (CD4+ Tmem), CD4+CD45RA+ T cells (CD4+ Tnaive), CD4+Tbet+CD45RA−T cells (Th1), CD4+Tbet+CD45RA+ T cells, CD25+FoxP3+CD4+ T cells (Tregs), CD8+CD45RA−T cells (CD8+ Tmem), CD8+CD45RA+ T cells (CD8+

Tnaive), CD8+Tbet+CD45RA− T cells, CD8+Tbet+CD45RA+ T cells, TCRγδ T cells, CD14+CD16− classical monocytes (cMCs), CD14−CD16+ non-classical monocytes (ncMCs), CD14+CD16+ intermediate monocytes (intMCs), monocytic myeloid-derived suppressor cells (M-MDSCs), myeloid dendritic cells (mDCs), and plasmacytoid dendritic cells (pDCs).

The signal intensity of 11 functional markers (pSTAT1, pSTAT3, pSTAT5, pSTAT6, pCREB, pMAPKAPK(pMK2), pERK, prpS6, pP38, and pNF- B, and total I B) for each cell type were simultaneously extracted. Two groupsκ κ of immune features were derived: endogenous signaling immune features and intracellular signaling response features. Endogenous intracellular signaling activities were derived from analyzing unstimulated blood samples. The arcsinh transformation of the median signal intensities of each signaling protein, per cell type, was used to calculate signaling immune features. Intracellular signaling response features were calculated as the difference in median signal intensity of each signaling protein (after arcsinh transformation) between stimulated and unstimulated samples.

Modeling and Analysis

Multivariate modeling of clock of postoperative recovery

A supervised prediction algorithm was developed to take a collection of physical activity representations (or features) and predict a patient’s trajectory of physical recovery from surgery. The presented approach extracts such activity and sleep monitoring features from a tridimensional accelerometer equipped in ActiGraph smart watches. These representations (e.g., step count, metabolic equivalent of task, onset of sleep and sleep efficiency) were computed according to state-of-the-art methods on unsupervised human activity learning and signal processing (see the Supplementary Table 1 for the complete list of features). The set of such computed representations for patient p are referred as feature matrix

F

_mn^(p)

=[f

_ij

]

for all i∈¿ and j∈¿ where m and n are the number of postoperative days and the number of features, respectively. Similarly, a baseline (preoperative) feature matrix for each patient is computed from five-day preoperative data (denoted by

B

^(p)

=[ b

_ij

]

for all i∈¿ and j∈¿ where k and n are the length of preoperative days and the number of features, respectively).

B

^(p) is used to compute a standardized transformation that independently scales feature columns to zero mean and unit variance:

f

_ij

:=(f

_ij

−μ

_j

)/ σ

_j where μ_j and σ_j are the mean and variance of feature j in

B

^{(p) . The}

computed transformation is then applied to

F

^(p) . Such transformation is commonly carried out for many machine learning estimators to reduce the unwanted effect of outliers or/and features with larger values on optimization—by having the training data look more similar to a standard Gaussian distribution.

The personalized prediction model

h

^(p) accepts an instance (i.e., computed features associated with sensor readings of patient p from a single day, denoted by

x

(ip)

= [ f

^ij

]

1×n

:∀ j ∈ ¿

and predicts the number of days passed since surgery (

^ y

^(p)_i ). This model can be formulated and optimized as a regression problem. As such, a number of classes of regression methods such as Generalized Linear Models, Decision Trees, Nearest Neighbors, and Ensemble Methods have been explored. Random Forest regression was chosen

as it benefits from generalizability of ensemble methods and flexibility of decision trees. Random Forest regression implements a bootstrap aggregation technique where it trains multiple decision tree models on random sampling with replacement of the training data (bootstrap) followed by aggregating the prediction of base models by averaging. As a result, it benefits the non-linearity of decision tree models while addressing the high variance of decision trees—boosting generalizability.

Quantifying patient-specific recovery surrogates from patients’ clock of postoperative recovery

The proposed surgical recovery surrogate vector represents an objective quantification of patient’s rate of recovery by transformation of the high dimensional time-series activity and sleep feature matrix to one dimensional ‘time to recovery’ surrogate vector. Each time-series activity feature when plotted against the baseline (preoperative) values visualizes a patient’s recovery trend over time with respect to its baseline (Figure 2C). Intuitively, one can define the recovery surrogate (denoted by

[ τ

j

]

1× n

: ∀ j∈ ¿

) as a vector of time after surgery when the value of clock of recovery returns to its prediction for baseline. Specifically, it is the first day since surgery (zero-based) where its corresponding feature value is at least as large as the Q₁ (or Q₃ if f>0 ) of baseline feature b_j . Simplistically, one could utilize the minimum (or maximum if f>0 ) of f_j as the threshold to returning to baseline status, however, the first quartile was picked to ensure robustness against any possible outlier in baseline feature b_j . For each patient, the recovery surrogate τ_j splits the postoperative timeline into two periods: day 0 to day

(τ

_j

−1 )

, when the patient is still in recovery; and day τ_j and after when the patient has recovered to his/her preoperative physical baseline as represented by j_th physical attribute. Given a patient p and its corresponding

B

^{(p) and}

F

^(p) for baseline and postoperative accelerometry feature matrices, respectively, the recovery surrogate per feature j (

τ

⁽_j^p) ) is formulated as follows. Assume f ≤0 without loss of generality.

τ

⁽_j^p)

=min

i∈¿

{ λ (i , p ) ,m−1 } s . t λ (i, p)= { i f m−1+ϵ

^ij(p)

≥Q

¹

o . w . (b

^(jp)

)

Multivariate modeling of recovery surrogates using Canonical Correlation Analysis

A CCA model was developed to investigate a linear association between patients’ recovery surrogate vector and immune and proteomics networks. Taking the correlated immune network, as an example, for a set of vectors

X =( X

1

, X

2

, …, X

p

)

of all endogenous signaling immune features and intracellular signaling response features in a training cohort of size p , and a set of corresponding recovery surrogate vectors

Y =(Y

₁

,Y

₂

, … ,Y

_p

)

, the model aims to maximize the coefficient of correlation between a linear combination of each set. This is carried out by finding a pair of canonical components

(C C

_X

, C C

_Y

)

such that C C_X=a₁X₁+a₂X₂+…+a_pX_p and C C_Y=a₁Y₁+a₂Y₂+…+a_pY_p .

Performance evaluation

To ensure independence between the observations in training and test, a repeated 10-fold cross validation has been applied to validate the random forest regression model. This approach makes certain that the models have been independently trained and tested on all patients without utilizing unseen patient data from the test splits during training. The 10-fold cross validation has been repeated 30 times using different random seed numbers to randomly create various splits of training and test data. The aggregated prediction for each input reported the median estimated time since surgery from the patient’s accumulated prediction vector. The prediction vector contained 30 predictions for each input (trained in 30 different folds, all blinded to the corresponding input). The Spearman’s correlation coefficient and p value between the aggregated predictions and the ground-truth values were reported (Supplementary Figure 3). The CCA model is validated using leave-one-patient-out cross validation to examine the generalizability of model prediction on an unseen patient while maximizing the size of training cohort in each iteration.

Supplementary Figure 1. CONSORT chart.

Supplementary Figure 2. Gating strategy.

Two-dimensional dot plots depicted for a representative sample. Gating was performed using Cell Engine (https://cellengine.com). Live singlet leukocytes were derived from DNA+ CD235ab– CD61– cPARP– events, which are split into mononuclear cells (CD45+ CD66– ) and neutrophils (CD45+ CD66+ ). Initially, adaptive cell populations are gated from the mononuclear cells (CD19+CD3– B cells and CD19– CD3+ T cells). The CD19– CD3– population is further split into CD7+ NK cells (CD56loCD16+ and CD56hiCD16–, respectively) and CD14+ CD16– classical monocytes, CD14+ CD16+

intermediate monocytes, and CD14– CD16+ non-classical monocytes. Dendritic cell populations are defined as CD14–

CD16– HLA-DR+ , and subdivided into CD11c+ mDC, and CD123+ pDC. A subpopulation of myeloid-derived suppressor cells (HLA-DRlo, CD11b+) is derived from classical monocytes. B cells are split into naïve (CD27– ) and memory (CD27+ ) cells. Transitional B cells (CD38+, CD24+ ) are derived from naïve cells and plasma (CD38+, CD27- ), and memory B cells (CD38- ;CD27+ ) from memory B cells. T cells are divided into TCRgd+ , CD4+, and CD8+ subpopulations. Among the CD4+

T cells, T helper 1 (Tbet+ ), regulatory T (FoxP3+ CD25+ ), naïve (CD45RA+ ), and memory (CD45RA– ) cells are identified.

Among the CD8+ T cells, naïve (CD45RA+ ), and memory (CD45RA– ) cells are identified. Cell types included in analysis are labeled in blue.

Supplementary Figure 3. Prediction of time since surgery.

A supervised prediction model using Random Forest (RF) regression was independently trained on the interconnected temporal graph of medium to high-level patient accelerometry attributes of each patient. The target variable to be predicted was time since surgery given the corresponding daily feature vector. The repeated 10-fold cross-validation showed high performance of the patient-specific RF model with median prediction p-values of

5.9× 10

^{−7 and a}

median spearman correlation coefficient of 0.7 across all models.

Supplementary Figure 4. Performance comparison of the clock of recovery vs the univariate models.

The multivariate model predicting time since surgery (the clock of surgical recovery) is compared to baseline univariate models trained on individual medium to high-level measurements. These measurements can be directly computed from ActiGraph smartwatches. Two metrics are shown in this comparison: Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The y axes show the errors (units are days since surgery) for all cross validated patients. Note that the same training and validation pipeline is used for this comparison.

Supplementary Figure 5. Correlation between preoperative immune features and postoperative physical recovery. (A) The bipartite correlation graph (spearman’s correlation) of immune and accelerometry networks. The edges correspond to significant correlation (p<0.05) between cellular responses (right) and ‘time to return to baseline’ for each accelerometry feature (left). The node colors in the accelerometry network indicate association with a major physical recovery cluster. The node colors in the immune networks indicate association with ex-vivo stimulations. (B) Examples of relevant strong correlations (edges) from each physical activity clusters ‘Activity Capacity’ and ‘Moderate and Overall Activity’ with pSTAT1 response to INF- stimulation in NK cells and CD4T cells (immune features shown on theα right) are depicted. The regression plot (regression line and 90% confidence interval) demonstrates the associated of each immune feature with speed of physical recovery with respect to the depicted physical recovery cluster over all patients.

Supplementary Figure 6. Pre-operative physical activity levels are not associated with the speed of recovery.

Preoperative physical activity and sleep quality measurements were used to predict the recovery surrogates derived from the proposed clock of postoperative recovery. A Random Forest regression model was trained and tested using a 10-fold cross-validation scheme. The recovery surrogates are plotted versus the model prediction using the test-subjects that were not used in model training. The results (Spearman’s ρ=0.08,p value=0.57 ) indicate that the recovery surrogate is not dependent on pre-operative physical activity levels.

Accelerometry features Description Activity Cluster ID NSBO (No. of Sedentary Bouts occurring) Number of Sedentary Bouts occurring in this day 0

NSBS (No. of Sedentary Bouts starting) Number of Sedentary Bouts starting in this day 0 NSBE (No. of Sedentary Bouts ending) Number of Sedentary Bouts ending in this day 0 NSBrO (No. of Sedentary Breaks occurring) Number of Sedentary Breaks occurring in this day 0 NSBrS (Sedentary Breaks Starting) Number of Sedentary Breaks starting in this day 0 NSBrE (Sedentary Breaks Ending) Number of Sedentary Breaks ending in this day 0

Vig (Vigorous) Length of time in Vigorous in minutes 0

VVig (Very Vigorous) Length of time in Very Vigorous in minutes 0

(%Vig) % Vigorous Percentage of time in Vigorous 0

(%VVig) % Very Vigorous Percentage of time in Very Vigorous 0

S_NSP Number of sleep periods for this day 0

S_AVGlate Average latency (from in bed to sleep onset) 0

S_LSMinutes Longest sleep period in minutes 0

S_NAwake The number of different awakening episodes as scored by the

algorithm. 0

S_FI The Fragmentation Index (FI) is the percentage of one minute

periods of sleep vs. all periods of sleep during the sleep period. 0

kcals kcals during this day 1

AHK (Average hourly kcals) Avg kcals per hour during this day 1

METs MET rate for a period of time 1

Moder (Moderate) Length of time in Moderate in minutes 1

(%Moder) % Moderate Percentage of time in Moderate 1

TMVPA (Total MVPA) Total time in moderate to vigorous activities in minutes 1

%MVPA Percent of time in MVPA 1

AMVPAPH (Average MVPA per hour) Average amount of MVPA per hour in this day 1

Axis1C (Axis 1 Counts) Sum of counts for Axis 1 (Y-Axis) 1

Axis2C (Axis 2 Counts) Sum of counts for Axis 2 (X-Axis) 1

Axis3C (Axis 3 Counts) Sum of counts for Axis 3 (Z-Axis) 1

Axis1AC (Axis 1 Average Counts) Average of counts for Axis 1 (Y-Axis) 1 Axis2AC (Axis 2 Average Counts) Average of counts for Axis 2 (X-Axis) 1 Axis3AC (Axis 3 Average Counts) Average of counts for Axis 3 (Z-Axis) 1

Axis1CPM (Axis 1 CPM) Counts Per Minute for Axis 1 (Y-Axis) 1

Axis2CPM (Axis 2 CPM) Counts Per Minute for Axis 2 (X-Axis) 1

Axis3CPM (Axis 3 CPM) Counts Per Minute for Axis 3 (Z-Axis) 1

VMC (Vector Magnitude Counts) Vector Magnitude of all 3 Axis 1

VMAC (Vector Magnitude Average Counts) Average Vector Magnitude of all 3 Axis 1

VMCPM (Vector Magnitude CPM) Vector Magnitude Counts Per Minute 1

StepsC (Steps Counts) Sum of Step Counts 1

StepsAC (Steps Average Counts) Average Step Counts 1

StepsPM (Steps Per Minute) Steps Per Minute 1

NBO (No. of Bouts Occuring) Number of Freedson (1998) Bouts occurring in this day 2 NBS (No. of Bouts Starting) Number of Freedson (1998) Bouts starting in this day 2 NBE (No. of Bouts Ending) Number of Freedson (1998) Bouts starting in this day 2 TTBO (Total time of Bouts occurring) Total time of Freedson (1998) Bouts occurring in this day 2 TACBO (Total activity counts of Bouts

occurring) Total activity counts of Freedson (1998) Bouts occurring in this

day 2

Axis1MC (Axis 1 Max Counts) Maximum count value for Axis 1 (Y-Axis) 2 Axis2MC (Axis 2 Max Counts) Maximum count value for Axis 2 (X-Axis) 2 Axis3MC (Axis 3 Max Counts) Maximum count value for Axis 3 (Z-Axis) 2 VMMC (Vector Magnitude Max Counts) Maximum Vector Magnitude of all 3 Axis 2

StepsMC (Steps Max Counts Maximum Step Counts 2

S_Effic. Sleep Efficiency – Number of sleep minutes divided by the total

number of minutes the subject was in bed 2 TTSBO (Total Time of Sedentary Bouts) Total time of Sedentary Bouts occurring in this day 3

Sed (Sedentary) Length of time in Sedentary in minutes 3

%Sed (% Sedentary) Percentage of time in Sedentary 3

S_TSMinutes The total number of minutes scored as “asleep.” 3 S_ALAwake The average length, in minutes, of all awakening episodes. 3

S_ActC The total actigraphy counts summed together for the entire sleep

period. 3

S_MI Movement Index (MI) is the percentage of epochs with y-axis

counts greater than zero in the sleep period. 3

S_SF The sum of MI and FI. 3

S_WASO Wake after Sleep Onset (WASO) – The total number of minutes

the subject was awake after sleep onset occurred. 3

S_MinutesLS Minutes past this day at in bed timestamp 4

TTSBrO (Time in Sedentary Breaks) Total time of Sedentary Breaks occurring in this day 5

Light Length of time in Light in minutes 5

%Light Percentage of time in Light 5

Supplementary Table 1. Daily accelerometry features and their corresponding activity cluster ids (0=’Overall Sleep’, 1=’Moderate and Overall Activity’, 2=’Activity Capacity’, 3=Sleep Disruption and Sedentary Activity’’, 4=’Onset of Sleep’, and 5=’Light Activity)

Antibody Manufacturer Metal Isotope Clone Concentration Comment

Barcode1 Trace Sciences Pd 102 15µM Barcode

Barcode2 Trace Sciences Pd 104 15µM Barcode

Barcode3 Trace Sciences Pd 105 15µM Barcode

Barcode4 Trace Sciences Pd 106 15µM Barcode

Barcode5 Trace Sciences Pd 108 15µM Barcode

Barcode6 Trace Sciences Pd 110 15µM Barcode

CD235ab* Biolegend In 113 HIR2 1µg/mL Phenotype

CD61* BD In 113 VI-PL2 0.5µg/mL Phenotype

CD45 Biolegend In 115 HI30 1µg/mL Phenotype

CD66 BD La 139 CD66a-B1.1 0.5µg/mL Phenotype

CD7 BD Pr 141 M-T701 0.5µg/mL Phenotype

CD19 Biolegend Nd 142 HIB19 0.5µg/mL Phenotype

CD45RA Biolegend Nd 143 HI100 0.5µg/mL Phenotype

CD11b Biolegend Nd 144 ICRF44 2µg/mL Phenotype

CD4 Biolegend Nd 145 RPA-T4 2µg/mL Phenotype

CD8a Biolegend Nd 146 RPA-T8 1µg/mL Phenotype

CD11c Biolegend Sm 147 Bu15 1µg/mL Phenotype

CD123 Biolegend Nd 148 6H6 1µg/mL Phenotype

pCREB Cell Signaling Technology Sm 149 87G3 2µg/mL Function

pSTAT5 Cell Signaling Technology Nd 150 C11C5 4µg/mL Function

p38 BD Eu 151 36/p38 2µg/mL Function

TCRgd BD Sm 152 B1 4µg/mL Phenotype

pSTAT1 BD Eu 153 14/P-STAT1 1µg/mL Function

pSTAT3 Cell Signaling Technology Sm 154 M9C6 2µg/mL Function

pS6 Cell Signaling Technology Gd 155 D57.2.2E 2µg/mL Function

CD24 Biolegend Gd 156 ML5 2µg/mL Phenotype

CD38 Gd 157 2µg/mL

CD33 Biolegend Gd 158 WM53 2µg/mL Phenotype

pMAPKAPK2 Cell Signaling Technology Tb 159 27B7 1µg/mL Function

Tbet Thermo Fisher Gd 160 4B10 8µg/mL Phenotype

cPARP** BD Dy 161 F21-852 1µg/mL Phenotype

FoxP3 Thermo Fisher Dy 162 PCH101 8µg/mL Phenotype

IkB Cell Signaling Technology Dy 164 L35A5 8µg/mL Function

CD16 Biolegend Ho 165 3G8 1µg/mL Phenotype

pNFkB BD Er 166 K10-895.12.50 2µg/mL Function

pERK Cell Signaling Technology Er 167 D13.14.4E 4µg/mL Function

pSTAT6 Biolegend Er 168 A15137E 1µg/mL Function

CD25 Biolegend Tm 169 M-A251 2µg/mL Phenotype

CD3 Biolegend Er 170 UCHT1 1µg/mL Phenotype

CD27 BD Yb 171 M-T271 2µg/mL Phenotype

CD15 Biolegend Yb 172 W6D3 8µg/mL Phenotype

CCR2 Biolegend Yb 173 K036C2 2µg/mL Phenotype

HLA-DR Fluidigm Yb 174 L243 2µg/mL Phenotype

CD14 Fluidigm Lu 175 M5E2 2µg/mL Phenotype

CD56 BD Lu 176 NCAM16.2 1µg/mL Phenotype

DNA1*** Fluidigm Ir 191 50µM DNA

DNA2*** Fluidigm Ir 193 50µM DNA

Supplementary Table 2. Mass cytometry panel.

Mass cytometry staining panel used in the study. *Antibodies targeting CD235ab and CD61 were combined into the same channel in order to identify erythrocytes and platelets as a group for their removal (gating out) from subsequent analysis.

**D214-cleaved PARP (cPARP) was used to identify apoptotic and pre-apoptotic cells for their removal (gating out) from subsequent analysis. ***An iridium-based DNA intercalator was used to help distinguish intact cells from cellular debris.

Stimulation Cell type Marker CCA Weight

GM-CSF CD24posCD38posBtrans STAT3 0.0575

IFNa CD27posBmem IkB 0.0502

IFNa CD27posCD38negB IkB 0.0486

GM-CSF CD27posCD38posBplasma STAT5 0.0483

IFNa intMCs IkB 0.0461

IFNa CD27negBnaive IkB 0.0460

IFNa Bcells IkB 0.0455

IL 2, 4 and 6 MDSCs STAT3 0.0445

IFNa Tregs IkB 0.0436

Unstim CD56posCD16negNK pSTAT6 0.0436

PI CD27posCD38posBplasma IkB 0.0430

GM-CSF CD56posCD16negNK S6 0.0410

GM-CSF CD27posBmem IkB 0.0403

GM-CSF CD27posCD38posBplasma STAT3 0.0402

IL 2, 4 and 6 CD8Tmem STAT3 0.0398

LPS CD27posCD38posBplasma STAT5 0.0398

IFNa cMCs IkB 0.0391

PI CD8Tmem NFkB 0.0386

GM-CSF CD24posCD38posBtrans MAPKAPK2 0.0384

Unstim CD56loCD16posNK p38 -0.0379

IFNa CD27posCD38posBplasma ERK -0.0379

PI mDCs MAPKAPK2 -0.0381

IFNa pDCs STAT5 -0.0381

PI Gr p38 -0.0381

IFNa NKcells STAT5 -0.0381

IFNa CD27posCD38negB pSTAT6 -0.0382

Unstim CD27posCD38negB NFkB -0.0382

IFNa pDCs pSTAT6 -0.0383

GM-CSF ncMCs NFkB -0.0383

IFNa intMCs p38 -0.0384

IFNa CD27posBmem pSTAT6 -0.0384

IFNa Bcells NFkB -0.0384

IL 2, 4 and 6 CD4Tmem STAT1 -0.0384

PI CD27negBnaive STAT5 -0.0386

GM-CSF Bcells STAT3 -0.0386

IL 2, 4 and 6 mDCs ERK -0.0387

IFNa CD56loCD16posNK STAT5 -0.0388

GM-CSF intMCs STAT3 -0.0389

GM-CSF CD27posCD38negB NFkB -0.0392

IL 2, 4 and 6 CD8Tmem pSTAT6 -0.0393

IFNa CD27posBmem ERK -0.0394

IL 2, 4 and 6 CD4TbetTmem STAT1 -0.0395

PI CD8Tnaive STAT5 -0.0395

LPS Bcells STAT1 -0.0396

IL 2, 4 and 6 Gr p38 -0.0396

LPS CD4TbetTmem STAT5 -0.0396

IL 2, 4 and 6 Bcells CREB -0.0397

PI pDCs STAT3 -0.0398

Unstim CD27negBnaive p38 -0.0401

Unstim gdTcells p38 -0.0402

LPS Gr NFkB -0.0402

LPS CD56posCD16negNK pSTAT6 -0.0402

GM-CSF CD56posCD16negNK IkB -0.0403

GM-CSF CD27negBnaive NFkB -0.0404

GM-CSF gdTcells STAT1 -0.0405

IFNa cMCs ERK -0.0406

GM-CSF CD4TbetTnaive STAT1 -0.0410

IFNa mDCs p38 -0.0411

Unstim pDCs STAT3 -0.0412

IL 2, 4 and 6 CD45RAposTregs STAT1 -0.0413

IFNa CD27posCD38negB ERK -0.0413

Unstim CD27posCD38posBplasma p38 -0.0413

IL 2, 4 and 6 CD4TbetTnaive IkB -0.0414

IFNa MDSCs ERK -0.0414

Unstim CD4Tcells STAT3 -0.0415

Unstim NKcells STAT3 -0.0415

Unstim Tregs STAT3 -0.0418

IL 2, 4 and 6 cMCs p38 -0.0418

Unstim CD27negBnaive NFkB -0.0418

IL 2, 4 and 6 cMCs MAPKAPK2 -0.0420

IL 2, 4 and 6 CD8Tmem CREB -0.0421

Unstim CD45RAposTregs STAT3 -0.0421

IL 2, 4 and 6 pDCs pSTAT6 -0.0422

IL 2, 4 and 6 Tregs STAT1 -0.0422

IL 2, 4 and 6 MDSCs p38 -0.0423

IL 2, 4 and 6 CD4Tnaive STAT1 -0.0424

PI CD4Tnaive STAT5 -0.0424

IL 2, 4 and 6 MDSCs NFkB -0.0424

IL 2, 4 and 6 CD45RAnegTregs STAT1 -0.0424

IL 2, 4 and 6 CD27posBmem STAT3 -0.0428

IL 2, 4 and 6 CD56loCD16posNK STAT1 -0.0429

LPS CD27negBnaive ERK -0.0432

IFNa Bcells CREB -0.0434

IL 2, 4 and 6 CD4TbetTnaive pSTAT6 -0.0434

Unstim CD56loCD16posNK STAT3 -0.0435

GM-CSF CD56posCD16negNK pSTAT6 -0.0436

PI mDCs STAT1 -0.0439

LPS CD4TbetTnaive IkB -0.0441

IL 2, 4 and 6 CD27posCD38negB STAT3 -0.0442

Unstim CD8Tnaive p38 -0.0443

IFNa CD27posCD38negB p38 -0.0444

Unstim CD45RAnegTregs STAT3 -0.0446

IFNa cMCs pSTAT6 -0.0447

Unstim CD4Tmem STAT3 -0.0447

IFNa CD27negBnaive p38 -0.0448

GM-CSF CD27negBnaive p38 -0.0449

IFNa CD27posBmem p38 -0.0450

IFNa Bcells p38 -0.0450

PI mDCs ERK -0.0452

IFNa CD56loCD16posNK STAT3 -0.0453

PI CD4TbetTnaive STAT1 -0.0453

IL 2, 4 and 6 CD56loCD16posNK pSTAT6 -0.0454

Unstim pDCs p38 -0.0458

Unstim CD4Tnaive STAT3 -0.0466

IL 2, 4 and 6 NKcells pSTAT6 -0.0468

IFNa CD8Tcells STAT3 -0.0468

LPS CD8Tmem pSTAT6 -0.0469

Unstim intMCs STAT3 -0.0470

PI mDCs pSTAT6 -0.0470

IFNa MDSCs STAT1 -0.0474

GM-CSF CD27posBmem NFkB -0.0475

LPS intMCs STAT3 -0.0476

IFNa CD4TbetTmem STAT1 -0.0479

IL 2, 4 and 6 CD56loCD16posNK STAT3 -0.0480

IFNa CD8Tnaive STAT3 -0.0480

IL 2, 4 and 6 CD27posBmem STAT1 -0.0482

IFNa CD4TbetTmem STAT3 -0.0484

IL 2, 4 and 6 Bcells p38 -0.0485

IFNa CD8Tmem STAT1 -0.0485

PI CD45RAposTregs STAT5 -0.0489

IFNa CD27posBmem CREB -0.0491

IL 2, 4 and 6 intMCs p38 -0.0491

IFNa MDSCs pSTAT6 -0.0491

IFNa ncMCs STAT3 -0.0492

IL 2, 4 and 6 CD27negBnaive p38 -0.0495

IFNa NKcells STAT3 -0.0495

IL 2, 4 and 6 NKcells STAT3 -0.0496

IFNa intMCs STAT1 -0.0498

IFNa cMCs STAT1 -0.0500

IL 2, 4 and 6 Bcells pSTAT6 -0.0500

IL 2, 4 and 6 CD27posBmem CREB -0.0503

IL 2, 4 and 6 NKcells STAT1 -0.0507

IL 2, 4 and 6 CD27posCD38negB STAT1 -0.0507

GM-CSF Bcells p38 -0.0511

IFNa CD4TbetTnaive STAT3 -0.0514

GM-CSF Bcells NFkB -0.0515

IL 2, 4 and 6 Bcells ERK -0.0516

IFNa CD8Tnaive STAT1 -0.0521

IL 2, 4 and 6 CD27negBnaive ERK -0.0521

LPS CD4TbetTmem STAT3 -0.0527

IL 2, 4 and 6 pDCs STAT5 -0.0529

IFNa mDCs STAT1 -0.0529

IFNa CD27posCD38negB CREB -0.0533

PI Bcells STAT5 -0.0537

IFNa CD27negBnaive STAT1 -0.0538

IL 2, 4 and 6 CD27posCD38negB CREB -0.0546

Unstim CD24posCD38posBtrans NFkB -0.0550

GM-CSF CD27posBmem CREB -0.0551

IL 2, 4 and 6 intMCs MAPKAPK2 -0.0551

IFNa CD8Tcells STAT1 -0.0552

IL 2, 4 and 6 gdTcells ERK -0.0552

IFNa MDSCs NFkB -0.0556

Unstim mDCs STAT3 -0.0559

Unstim cMCs STAT3 -0.0561

IFNa CD27negBnaive ERK -0.0563

IFNa Bcells STAT1 -0.0563

Unstim CD56posCD16negNK p38 -0.0567

IFNa CD27negBnaive STAT3 -0.0573

IFNa CD27posBmem STAT3 -0.0575

IFNa CD27posBmem STAT1 -0.0576

IFNa CD56posCD16negNK STAT3 -0.0581

IFNa cMCs p38 -0.0589

IFNa Bcells STAT3 -0.0595

IFNa CD27posCD38negB STAT1 -0.0597

IL 2, 4 and 6 CD27posCD38negB ERK -0.0598

IFNa CD27posCD38negB STAT3 -0.0606

IFNa CD45RAnegTregs STAT1 -0.0607

IFNa Bcells ERK -0.0609

IL 2, 4 and 6 CD27posBmem ERK -0.0614

LPS pDCs STAT3 -0.0614

GM-CSF CD4TbetTnaive STAT3 -0.0618

IFNa MDSCs p38 -0.0619

IL 2, 4 and 6 CD8Tmem ERK -0.0644

IFNa CD4Tmem STAT1 -0.0646

IFNa CD45RAposTregs STAT1 -0.0646

Unstim MDSCs STAT3 -0.0647

IFNa CD4Tnaive STAT1 -0.0650

IFNa Tregs STAT1 -0.0658

GM-CSF CD27posCD38negB CREB -0.0673

IFNa CD4Tcells STAT1 -0.0676

IFNa CD56posCD16negNK STAT1 -0.0683

IFNa CD56loCD16posNK STAT1 -0.0699

IFNa NKcells STAT1 -0.0717

IFNa CD24posCD38posBtrans pSTAT6 -0.0726

Supplementary Table 3. Coefficients of mass cytometry CCA model.

Coefficients (weights) of the Canonical Correlation Analysis (CCA) model trained on mass cytometry features predicting surrogates of postoperative recovery. The 10 percent features in terms of significance to the predictive model are listed.

The positive (or negative) weights indicate positive correlation with larger (or smaller) values of activity surrogates.

References

[1] B. Gaudillière et al., “Clinical recovery from surgery correlates with single-cell immune signatures,” Sci. Transl.

Med., vol. 6, no. 255, pp. 255ra131-255ra131, Sep. 2014, doi: 10.1126/scitranslmed.3009701.

[2] J. S. Paddison, R. J. Booth, A. G. Hill, and L. D. Cameron, “Comprehensive assessment of peri-operative fatigue:

development of the Identity-Consequence Fatigue Scale,” J. Psychosom. Res., vol. 60, no. 6, pp. 615–622, Jun. 2006, doi: 10.1016/j.jpsychores.2005.08.008.

[3] J. S. Paddison, T. Sammour, A. Kahokehr, K. Zargar-Shoshtari, and A. G. Hill, “Development and validation of the surgical recovery scale (SRS),” J. Surg. Res., 2011, doi: 10.1016/j.jss.2010.12.043.

[4] J. Woodfield et al., “Protocol, and practical challenges, for a randomised controlled trial comparing the impact of high intensity interval training against standard care before major abdominal surgery: study protocol for a randomised controlled trial,” Trials, vol. 19, no. 1, p. 331, Dec. 2018, doi: 10.1186/s13063-018-2701-9.

[5] J. A. Duffield et al., “Intraperitoneal local anesthetic instillation and postoperative infusion improves functional recovery following colectomy: A randomized controlled trial,” Dis. Colon Rectum, 2018, doi:

10.1097/DCR.0000000000001177.

[6] P. P. Singh, S. Srinivasa, D. P. Lemanu, A. A. Kahokehr, and A. G. Hill, “The Surgical Recovery Score correlates with the development of complications following elective colectomy,” J. Surg. Res., 2013, doi: 10.1016/j.jss.2012.12.005.

[7] J. M. McDonnell, D. P. Ahern, T. D. Ross, D. Gibbons, K. A. Synnott, and J. S. Butler, “The efficacy of remote virtual care in comparison to traditional clinical visits for elective orthopaedic patients: A meta-analysis of prospective randomised controlled trials,” Surg., Mar. 2021, doi: 10.1016/j.surge.2021.02.008.

[8] F. Leiss et al., “Excellent Functional Outcome and Quality of Life after Primary Cementless Total Hip Arthroplasty (THA) Using an Enhanced Recovery Setup,” J. Clin. Med., 2021, doi: 10.3390/jcm10040621.

[9] T. Wang, Y. Zhou, X. Li, S. Gao, and Q. Yang, “Comparison of postoperative effectiveness of less invasive short external rotator sparing approach versus standard posterior approach for total hip arthroplasty,” J. Orthop. Surg.

Res., vol. 16, no. 1, p. 46, Dec. 2021, doi: 10.1186/s13018-020-02188-2.

[10] L. Dumenci et al., “Model-based pain and function outcome trajectory types for patients undergoing knee arthroplasty: a secondary analysis from a randomized clinical trial,” Osteoarthr. Cartil., 2019, doi:

10.1016/j.joca.2019.01.004.

[11] E. A. Brembo, H. Kapstad, S. Van Dulmen, and H. Eide, “Role of self-efficacy and social support in short-term recovery after total hip replacement: a prospective cohort study,” Health Qual. Life Outcomes, vol. 15, no. 1, p. 68, Jan. 2017, doi: 10.1186/s12955-017-0649-1.

[12] G. K. Fragiadakis et al., “Patient-specific Immune States before Surgery Are Strong Correlates of Surgical Recovery.,” Anesthesiology, vol. 123, no. 6, pp. 1241–55, Dec. 2015, doi: 10.1097/ALN.0000000000000887.

[13] E. A. Ganio et al., “Preferential inhibition of adaptive immune system dynamics by glucocorticoids in patients after acute surgical trauma,” Nat. Commun., vol. 11, no. 1, pp. 1–12, Dec. 2020, doi: 10.1038/s41467-020-17565-y.