Contents lists available at SciVerse ScienceDirect

Journal of Veterinary Behavior 8 (2013) 362e366

Journal of Veterinary Behavior

journal homepage: www.journalvetbehavior .com

Research

ROC analysis of prepartum feeding time can accurately predict postpartummetritis development in HF crossbred cows

Tapas Kumar Patbandha a,#, Tushar Kumar Mohanty a,*, Siddhartha Shankar Layek a,Arumugam Kumaresan a, Suresh C. Kantwa a, R. Malhotra b, A.P. Ruhil c, Shiv Prasad a

aCattle Yard, Livestock Production and Management Section, National Dairy Research Institute, Karnal, Haryana, IndiabDESM Division, National Dairy Research Institute, Karnal, Haryana, IndiacASRB, Krishi Bhawan, Pusa, ICAR, New Delhi, India

a r t i c l e i n f o

Article history:Received 25 June 2012Received in revised form14 April 2013Accepted 1 May 2013Available online 17 June 2013

Keywords:ROC analysisfeeding timemetritisthreshold valuerelative risk

* Address for reprint requests and correspondenCattle Yard, Livestock Production and Management SeInstitute, Karnal 132001, Haryana, India; Tel: þ91-921

E-mail address: mohanty.tushar@gmail.com (T.K.# Current address: Department of Livestock Product

of Veterinary Science and Animal Husbandry, JunaJunagadh 362001, Gujarat, India.

1558-7878/$ e see front matter � 2013 Elsevier Inc.http://dx.doi.org/10.1016/j.jveb.2013.05.001

a b s t r a c t

Receiver operating characteristic (ROC) analysis is a useful statistical tool to visualize, organize, and selectdiagnostic variables. Accuracy of the test is represented by the area under the ROC curve (AUC), sensi-tivity (Se), and specificity (Sp). In the present study, ROC analysis of prepartum feeding time (day �11 today �2) of 20 pluriparous Holstein-Friesian crossbred cows were recorded, and the cows were followedup for detecting postpartum metritis development if any. Prepartum daily feeding time data duringwhole study was grouped into 2 periods (P1 [day �11 to day �7] and P2 [day �6 to day �2]). The AUC offeeding time for the P1 and P2 periods was 0.64 (P ¼ 0.29) and 0.86 (P ¼ 0.006), respectively. Cowswith daily feeding times below the optimum threshold value of 284.5 min/d during P2 (Se ¼ 75% andSp ¼ 91.67%) had more chance of developing postpartum metritis compared with those with dailyfeeding time above the threshold value (positive likelihood ratio ¼ 9.0). Furthermore, relative risk (RR)analysis revealed that the risk of development of postpartum metritis increased as feeding timedecreased below the optimum threshold value during P2 (RR ¼ 5.571, P ¼ 0.004). Thus, the optimumthreshold value of feeding time during P2 may be used to predict risk of postpartum metritis, allowingfor preventive measures to be implemented in the early postpartum period to reduce the likelihood ofpostpartum complications.

� 2013 Elsevier Inc. All rights reserved.

Introduction

Metritis is an important postpartum health disorder, which af-fects reproductive performance and productivity of high-yieldingdairy cows and buffaloes. Heavy productivity loss because ofincreased days open, days dry, and calving interval along withsignificant reduction in lactation length has been reported incrossbred cows at National Dairy Research Institute, Karnal(Balasundaram et al., 2011; Sharma et al., 2011). Earlier studies onblood metabolites and feeding behavior revealed that these pa-rameters can be effectively used for early prediction of metritis

ce: Tushar Kumar Mohanty,ction, National Dairy Research5508002.Mohanty).ion and Management, Collegegadh Agricultural University,

All rights reserved.

(Urton et al., 2005; Huzzey et al., 2007; Ospina et al., 2010). Pre-partum feeding time below an optimum threshold value (75 min/d)was found to be highly correlated with increased incidence ofpostpartum metritis (Urton et al., 2005). The optimum thresholdvalue is defined as the threshold value that has maximum com-bined sensitivity (Se) and specificity (Sp) for predicting diseases(Urton et al., 2005; Ospina et al., 2010).

The common method used to find the optimum threshold valueis through receiver operating characteristic (ROC) analysis. It is asimple statistical tool used to characterize a variable in terms ofarea under the ROC curve (AUC) and provides optimum thresholdvalues along with the corresponding Se, Sp, and likelihood ratio(LR). The Se measures the proportion of sick and animals that areidentified by the diagnostic test as sick (true positives [TPs]),whereas Sp measures the proportion of healthy cows correctlyidentified by the test as healthy (true negatives [TNs]). The LR refersto the number of times a cow is more likely to develop a diseasewhen they are below a particular threshold value relative to if theywere above this threshold value. The assumption behind ROC

T.K. Patbandha et al. / Journal of Veterinary Behavior 8 (2013) 362e366 363

analysis is that a diagnostic variable (continuous classifier, i.e.,feeding time) has binary outcomes (metritis or normal).

The optimum threshold value for daily feeding time to predictensuing metritis in medium- to high-yielding crossbred cows is notavailable. Hence, the present study was conducted in Holstein-Friesian (HF) crossbred cows to find out a suitable optimumthreshold value for prepartum feeding time by ROC analysis forearly prediction of metritis along with its Se, Sp, and LRþ. Usingthese optimum threshold values, the relative risk (RR) of develop-ment of postpartum metritis was also estimated.

Materials and methods

Experimental animals and management

The present experiment was conducted at Cattle Yard, NationalDairy Research Institute, Karnal. Karnal is located on 29� 420 Nlatitude and 72� 020 E longitude at an altitude of 250 m above themean sea level. Twenty pluriparous dry pregnant HF crossbredcows (HF � Tharparkar) were based on normal body conditionscore (3.5-4.5 in 6-point scale; Prasad, 1994); lactation yield (3800-4500 L during 305 days) and body weight (450-550 kg) wereselected. Experimental cows were maintained in a loose housingsystem of management during December to March (2010-2011).The temperature during study period ranged from 2�C (minimum)to 30�C (maximum) and relative humidity from 35% (minimum) to96% (maximum). The cows were fed measured amount of concen-trate (20% crude protein and 70% total digestible nutrients) as perNational Research Council (2001) standards. The concentratemixture comprised of maize (33%), groundnut cake (21%), mustardcake (12%), wheat bran (20%), deoiled rice bran (11%), mineralmixture (2%), and common salt (1%). The concentrate mixture fedwas 1.5 kg/d/cow for body maintenance and extra concentrate of1.5 kg/cow 3 weeks before expected date of calving. In addition toconcentrate, ad libitum seasonal green fodder (berseem and oat)was provided 2-3 times a day. The dry matter (DM) percent ofconcentrate, berseem, and oat was approximately 90, 12, and 14%,respectively. Daily ration was given in the common manger locatedat one side of the loose house with adequate manger space (0.6 mper animal).

Monitoring of feeding time and metritis diagnosis

The cows were moved to an experimental paddock to monitorand record behavior 3 weeks before the expected date of calving.When a cow started showing physical signs of imminent calving(udder enlargement, milk letdown, and relaxation of sacrosciaticligament), she was shifted to the calving pen. Cows were markedwith hair dye on their back and sides for easy identification. Feedingbehavior (daily feeding time) was monitored continuously bymeans of 3 video cameras with infrared lighting arrangement andrecorded by an 8-channel digital video recorder (MPEG4/H.264; TSMicroTech, Inc, California, USA). A cowwas considered to be feedingwhen its head was in the manger with active ingestion. Dailyfeeding time datawere analyzed by replay of recorded datawith 2�to 32� speed depending on feeding activity pattern. Start and endpoints of feeding were noted to calculate duration of feeding time(difference between start and end points).

Cows were regularly monitored (3-4 days of interval) for healthproblems up to 21 days postpartum, and uterine discharge wascollected through rectovaginal method for metritis diagnosis. Theuterine discharge was aspirated from the uterus using sterile bluesheath (IMV Technologies, L’Aigle, France) and universal artificialinsemination gun and transferred to a clean test tube. Cows thatshowed fetid uterine discharge (purulent or redebrown) along

with high body temperature (>39.5�C) within 3 weeks postpartumwere diagnosed as metritic (Sheldon et al., 2006).

Statistical analysis

In the present study, owing to variation between expected andactual date of calving, we recorded complete daily feeding time datafrom day �11 to day �2 during prepartum period. The prepartumdaily feeding time data during the whole study were grouped into 2periods, that is, P1 (day �11 to day �7) and P2 (day �6 to day �2),and the cows were divided into 2 groups as normal and metriticafter diagnosis of metritis. Descriptive analysis was conducted tofind the mean, standard error, and range of daily feeding time.Feeding time, dependent variable, was analyzed by 2-way analysisof variance to evaluate the effect of health (metritis and normal)and periods (P1 and P2) on this measure. All pairwise differences inmean were compared by Tukey post hoc test. Different optimumthreshold values of feeding time were analyzed by ROC analysisusing SigmaPlot 11 software package (Systat software, Inc, Califor-nia, USA). Power analysis was conducted using MedCalc, version12.3 software package (MedCalc Software, Ostend, Belgium). TheROC analysis produces AUC, a 2-dimensional graph (Se and 1 � Spplotted in Y-axis and X-axis, respectively, for the range of potentialoptimum threshold values), which gives the accuracy. Accuracy ofdiagnostic test was interpreted based on the AUC and said to benoninformative if AUC is 0.5, accurate if AUC is 0.5-0.7, very accurateif AUC is 0.7-0.9, highly accurate if AUC is 0.9-1.0, and finally perfectif AUC is 1 (Swets, 1988).

The Se values depend on the number of TP and false negative(FN) cases, whereas Sp values depend on the number of TN andfalse positive (FP) cases. Furthermore, the Se and Sp together definethe positive LR (LRþ). The TP indicates the number of metritic cowshaving prepartum feeding time below the threshold value, and FPindicates the number of metritic cows having prepartum feedingtime above the threshold value. The TN indicates the number ofnormal cows having prepartum feeding time above the thresholdvalue, and FN indicates the number of normal cows having pre-partum feeding time below the threshold value (Lalkhen andMcCluskey, 2008). The Se was defined as the proportion of cowswithmetritis that had feeding time below the threshold value ([TP]/[TP þ FN]). The Sp was the proportion of normal cows that hadfeeding time above the threshold value ([TN]/[TN þ FP]). The LRþwas the number of times a cowwas more likely to develop metritisbelow a particular threshold value (LRþ ¼ Se/[1 � Sp]) (Greineret al., 2000; Akobeng, 2007; Fawcett, 2006; Faustini et al., 2007;Lalkhen and McCluskey, 2008).

The optimum threshold value of feeding time was used todichotomize cattle into 2 health categories (feeding time aboveeach threshold value indicated “0” value [healthy] and below eachthreshold value indicated “1” [metritic]), and a simple 2 � 2 con-tingency table was constructed to analyze RR of metritis (Simon,2001; Cummings, 2009). The RR is the ratio of proportion of cowsthat had feeding time below the threshold value developedmetritisto the proportion of cows that had feeding time above the thresholdvalue developed metritis ({TP/[TPþ FN]}/{FP/[FPþTN]}).

Results

Feeding time

Among the 20 experimental cows, 8 cows developed metritisduring postpartum period. There was a significant difference (P <

0.001) in prepartum feeding time between normal and metriticcows during both the periods studied (Table 1). Similarly, thefeeding time of cows irrespective of the health status differed

Table 2Receiver operating characteristic analysis for optimum threshold of feeding timeduring period P2 (day �6 to day �2)

Feeding time thresholdvalues (min/d)

Se (%) 95% CI for Se Sp (%) 95% CI for Sp LRþ

274.7 50.0 15.70-84.30 91.67 61.52-99.79 6.00279.3 62.5 24.49-91.48 91.67 61.52-99.79 7.50284.5* 75.0 34.91-96.81 91.67 61.52-99.79 9.00287.1 75.0 34.91-96.81 83.33 51.59-97.91 4.49288.4 75.0 34.91-96.81 66.67 34.89-90.08 2.25295.2 75.0 34.91-96.81 58.33 27.67-84.83 1.79302.6 87.5 47.35-99.68 58.33 27.67-84.83 2.09304.0 87.5 47.35-99.68 50.0 21.09-78.91 1.75

Se, sensitivity; Sp, specificity; CI, confidence interval; LRþ, likelihood ratio positive.* Optimum threshold value based on maximum Se and Sp.

Table 1Least square mean of feeding time (min/d) of metritic and normal cows during P1(day �11 to day �7) and P2 (day �6 to day �2)

Period Feeding time (mean � standard error)

Metritis (n ¼ 8) Normal (n ¼ 12)

P1 (day �11 to day �7) 361.95 � 65.80Aa 399.31 � 63.66Ab

P2 (day �6 to day �2) 267.47 � 45.72Ba 318.28 � 51.16Bb

Means with different superscripts in a column (A, B) and row (a, b) differ signifi-cantly (P < 0.001).

T.K. Patbandha et al. / Journal of Veterinary Behavior 8 (2013) 362e366364

significantly (P < 0.001) between the time periods (Table 1). Met-ritic cows spent less time feeding (P < 0.001) than normal cowsduring both the periods, that is, P1 (day �11 to day �7) and P2(day �6 to day �2).

Critical threshold values and positive LRs

The AUC of feeding timewas 0.64 (P¼ 0.29) and 0.86 (P¼ 0.006)during P1 and P2, respectively (Figure 1). Critical threshold valuesalong with corresponding Se and Sp of prepartum feeding time forpredicting postpartummetritis during period P2 (day�6 to day�2)are presented in Table 2. The optimum threshold value duringperiod P2 was 284.5 min/d (Figure 2). The dot histograms of dailyfeeding time of normal and metritic cows during P1 and P2 arepresented in Figure 3. The dot histogram reflected that feeding timeduring P1 ranged from minimum 275.5 min/d to maximum499.4 min/d and during P2 from 220.0 min/d (minimum) to380.2 min/d (maximum). For different threshold values, the LRþvalues are given in Table 2 (P2). The LRþ value for optimumthreshold value during P2 was 9.0. Retrospective power analysisbased on AUC during P1 (AUC ¼ 0.64) revealed that total samplesize required for 80% power with a ¼ 0.05 during P1 was 264.

Relative risk

The RR values of different threshold values of daily feeding timeduring P2 for development of postpartummetritis are presented in

Figure 1. Area under the receiver operating characteristic curve (AUC) during P1(day �11 to day �7) and P2 (day �6 to day �2).

Table 3. In addition to the optimum threshold value of daily feedingtime identified through AUC analysis (i.e., 284.5 minutes), thethreshold values 279.3 and 287.1 min/d were also significantlyassociated with the risk of development of metritis (Table 3).

Discussion

The prepartum daily feeding time of normal and metritic cowswas compared to see if feeding time can be used to predict ensuingmetritis. In the present study, we observed a significant differencein prepartum feeding time between normal and metritic cows.Lower daily feeding time in metritic cows observed in the presentstudy is similar to those reported by Urton et al. (2005) and Huzzeyet al. (2007), who observed less prepartum daily feeding time inmetritic cows. Although we observed a lower feeding time inmetritic cows compared with normal cows, the daily feeding timein both the groups (normal andmetritic) was higher comparedwiththose reported earlier (Urton et al., 2005; Huzzey et al., 2007).Higher feeding time in our study might be because of the type ofdaily ration offered to the cows, that is, more roughage in the totaldiet instead of total mixed ration (TMR), which increases the dailyfeeding time. The green roughage, berseem, and oat contain lessDM, that is, 12%-14%, so to meet the required amount of DM intake,the cows might have spent more time at the manger. Huzzey et al.(2007) reported a gradual reduction of feeding time as the cowapproached toward parturition, which was true in our study also.The increase in size of the growing fetus toward parturition reducesrumen space by putting pressure on it and may reduce DM intakeleading to reduced feeding time. Lower DM intake before calvingcan worsen negative energy balance in transition cows, which mayin turn suppress immune status and increases susceptibility topostpartum uterine infection (Hammon et al., 2006).

Figure 2. Feeding time optimum threshold value (min/d) during P2 (day �6 today �2).

Figure 3. Dot histogram of feeding time of normal (n ¼ 12) and metritic (n ¼ 8) cowsduring P1 (day �11 to day �7) and P2 (day �6 to day �2).

T.K. Patbandha et al. / Journal of Veterinary Behavior 8 (2013) 362e366 365

The accuracy of any diagnostic test is reflected by the value ofAUC in ROC analysis. Swets (1988) reported that the AUC would beinformative to interpret the accuracy of diagnostic test only when itis more than 0.5 and perfect when it is 1. AUC analysis revealed thatonly a feeding time threshold during P2 could be used to predictmetritis. Low sample size in combination with the large range infeeding times during P1 might have explained why a significantfeeding time threshold could not be identified during P1. In thepresent study, although the sample size was same for both theperiods, a retrospective power analysis revealed that a significantlygreater sample size was required during P1. The daily feeding timeduring P1 was distributed in larger range compared with P2 aspresented by the dot histogram, which may lead to more FP and FNcases. During P1, the magnitude of the difference in feeding timebetween health groups was smaller than during P2, which isanother reason why it was probably harder to identify an optimalthreshold value to predict metritis.

The threshold value having highest combining Se and Sp iscalled an optimum threshold value, which is suitable for predictingensuing diseases. In the present study, we observed higher opti-mum threshold values for feeding time than previous study (Urtonet al., 2005) for metritis diagnosis. Urton et al. (2005) reported acritical threshold value of 75 min/d feeding time during peripartumperiod in Holstein cows fed with TMR. The higher optimumthreshold value observed in the present study may be because ofthe feeding of cows with more roughage in the total ration. The Seand Sp of optimum threshold value during P2 were 75% versus91.67%, respectively. The Se of 75% indicates that optimumthreshold value of daily feeding time 284.5 min/d during P2 couldcorrectly identify 75% of metritic cows (TP cases) but, 25% cases of

Table 3Association between feeding time during P2 (day �6 to day �2) and RR of devel-opment of metritis

Feeding time thresholdvalues (min/d)

RR 95% CI for RR P value

274.7 3.000 1.164-7.732 0.109279.3 3.889 1.341-11.281 0.018284.5* 5.571 1.503-20.653 0.004287.1 4.500 1.194-16.962 0.019288.4 3.000 0.786-11.445 0.170295.2 2.455 0.645-9.337 0.197302.6 4.667 0.702-31.036 0.070

CI, confidence interval; RR, relative risk.* Optimum threshold value based on maximum sensitivity and specificity.

metritis may go unnoticed (FN cases). On the other hand, 91.67% ofSp indicated that optimum threshold value of daily feeding timeduring P2 could correctly identify approximately 92% normal cows(TN cases), but 8% normal cows may be incorrectly identified asmetritic (FP cases). We observed a comparable Se and higher Spcompared with previous study (Urton et al., 2005). Depending onthe degree of severity of uterine infections, Urton et al. (2005)observed Se that ranged from 71% to 89% and Sp that ranged from62% to 77% in Holstein cows.

LRþ is the number of times a cow is more likely to developmetritis below a particular threshold value of prepartum feedingtime. The LRþ value for optimum threshold value of feeding timeduring P2 indicates that the cows with daily feeding time below theoptimum threshold value of 284.5 min/d were 9 times more likelyto develop postpartum metritis. Ospina et al. (2010) reported arelation between energy metabolites and LRþ for prediction ofmetritis. The cows with elevated prepartum nonesterified fatty acidabove optimum threshold value were 1.9 times more likely todevelop postpartum metritis (Ospina et al., 2010). RR is anothermeasure used to describe the risk of developing a disease in thefuture. Although we observed a significant association betweendifferent threshold values of prepartum dairy feeding time andpostpartum metritis (Table 3), the RR value was highest for theoptimum threshold value (i.e., 284.5 minutes). The present studyrevealed that the risk of development of metritis increased by5.5 times as the daily feeding time of cow decreased below284.5 min/d during P2. Although the sample size in the presentstudy is small, it provides information about the association ofdecreased prepartum daily feeding time with increased incidenceof metritis. Furthermore, detailed study need to be conducted withoptimum sample size to find a farm-specific optimum thresholdvalue for early identification of ensuing metritis of HF crossbredcows.

Conclusions

The optimum feeding time threshold during P2 (day �6 today �2) for accurately predicting risk of postpartum metritis incrossbred cows is 284.5 min/d. Cows with feeding times below thisthreshold were most likely to develop metritis. This thresholdyields the highest combined Se and Sp and also is associated withthe greatest RR and LR estimates.

Acknowledgments

The authors are highly thankful to director and vice-chancellorof National Dairy Research Institute, Karnal, for providingresearch facilities. The first author is a recipient of Junior ResearchFellowship from Indian Council of Agricultural Research, NewDelhi,and the work was funded by National Agricultural InnovationProject (NAIP/C4/C2008). We are also thankful to Dr. SubhashChand for his technical assistance during the study.

References

Akobeng, A.K., 2007. Understanding diagnostic tests 1: sensitivity, specificity andpredictive values. Acta Paediat. 96, 338e341.

Balasundaram, B., Gupta, A.K., Dongre, V.B., Mohanty, T.K., Sharma, P.C., Khate, K.,Singh, R.K., 2011. Influence of genetic and non-genetic factors on incidence ofcalving abnormalities in Karan Fries cows. Ind. J. Anim. Res. 45, 26e31.

Cummings, P., 2009. The relative merits of risk ratios and odds ratios. Arch. Pediatr.Adolesc. Med. 163, 438e445.

Faustini, M., Battocchio, M., Vigo, D., Prandi, A., Veronesi, M.C., Comin, A., Cairoli, F.,2007. Pregnancy diagnosis in dairy cows by whey progesterone analysis: an ROCapproach. Theriogenology 67, 1386e1392.

Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recognit. Lett. 27, 861e874.

T.K. Patbandha et al. / Journal of Veterinary Behavior 8 (2013) 362e366366

Greiner, M., Pfeiffer, D., Smith, R.D., 2000. Principles and practical application of thereceiver-operating characteristic analysis for diagnostic tests. Prev. Vet. Med. 45,23e41.

Hammon, D.S., Evjen, I.M., Dhiman, T.R., Goff, J.P., Walters, J.L., 2006. Neutrophilfunction and energy status in Holstein cows with uterine health disorders. Vet.Immunol. Immunopathol. 113, 21e29.

Huzzey, J.M., Veira, D.M., Weary, D.M., Von Keyserlingk, M.A.G., 2007. Prepartumbehavior and dry matter intake identify dairy cows at risk for metritis. J. DairySci. 90, 3220e3233.

Lalkhen, A.G., McCluskey, A., 2008. Clinical tests: sensitivity and specificity. Contin.Educ. Anaesth. Crit. Care Pain 8, 221e223.

National Research Council, 2001. Nutrient Requirements of Dairy Cattle, 7th ed.National Academy Press, Washington, DC.

Ospina, P.A., Nydam, D.V., Stokol, T., Overton, T.R., 2010. Evaluation of nonesterifiedfatty acids and b-hydroxybutyrate in transition dairy cattle in the north-eastern

United States: critical thresholds for prediction of clinical diseases. J. Dairy Sci.93, 546e554.

Prasad, S., 1994. Body condition scoring and feeding management in relation toproduction performance of cross bred dairy cattle [PhD thesis]. National DairyResearch Institute, Karnal, India.

Sharma, P.C., Gupta, A.K., Mohanty, T.K., Chakravarty, A.K., Singh, A., Chand, S., 2011.Uterine and ovarian health traits in Karan Fries cows. Ind. Vet. J. 88, 83e85.

Sheldon, I.M., Lewis, G.S., LeBlanc, S., Gilbert, R.O., 2006. Defining postpartumuterine disease in cattle. Theriogenology 65, 1516e1530.

Simon, S.D., 2001. Understanding the odds ratio and the relative risk. J. Androl. 22,533e536.

Swets, J.A., 1988. Measuring the accuracy of diagnostic systems. Science 240, 1285e1293.

Urton, G., Von Keyserlingk, M.A.G., Weary, D.M., 2005. Feeding behavior identifiesdairy cows at risk for metritis. J. Dairy Sci. 88, 2843e2849.