The objective of this study was to evaluate the feasibility of using multiple 3-dimensional accelerometers to estimated individual dry matter intake (DMI) of lactating dairy cows. Twenty-four Holstein cows in late lactation were assigned into 2 groups, a calibration group (n = 12) and a validation group (n = 12). All cows were fitted with 3 sensors that recorded 3-dimensional acceleration (i.e., x, y, and z) at 10-s intervals, 1 on the lateral side of the left hind leg and 2 attached directly to a halter over the nose and jaw area on the left side. Then, 3 accelerations were generated from each accelerometer (e.g., Leg-X, Leg-Y, and Leg-Z). Six new variables were created based on the change in acceleration in the nose and jaw accelerometers between 2 consecutive time points (e.g., LagJaw-X). For both groups (i.e., calibration and validation), cows were continuously video recorded while data on acceleration and intake of total mixed ration were collected for 10 consecutive days. Cows were fed once daily using an individual gate system, and individual refusals were recorded next day before morning feeding. Cows were fed a common lactating cow diet (17.9% crude protein; 1.70 Mcal/kg of dry matter). In the calibration group, individual eating bouts were obtained based on video recordings and merged with the corresponding accelerometer data. Then, a stepwise regression analysis was conducted using the REG procedure of SAS (SAS Institute, Cary, NC) to determine the ranges in acceleration that accounted for the highest variation in DMI (highest R2) in each acceleration variable. All 32,767 potential acceleration combinations were tested in the validation group using the acceleration ranges predetermined in the calibration group. The CORR procedure of SAS was used to test the Pearson correlation coefficient (r) between the type of DMI [i.e., based on accelerations (DMIaccel) or actual DMI (DMIactual)]. The MIXED procedure of SAS was used to perform a repeated-measures analysis with type (DMIaccel vs. DMIactual), day, and their interaction (T × D) in the model. From this analysis, 8 candidate acceleration models were selected based on high r and similarity (P > 0.15) in terms of T and T × D between DMIaccel and DMIactual. A simulated effect on DMIactual was artificially created in the validation group by dividing this group (n = 12) into high and low intake cows (n = 6/group; DMI of 24.1 vs. 18.7 kg/d), and the candidate models were tested to determine whether they were sensitive enough to detect this effect. From these candidate models, AEN (Leg-X + Jaw-Z + LagJaw-Z) showed a weak correlation (r = 0.36) between DMIaccel and DMIactual, but DMIaccel and DMIactual were highly similar (21.2 vs. 21.4 kg/d of DMI). In addition, this was the only model that could detect the simulated effect on DMIactual (22.1 vs. 20.3 kg/d of DMI) in the validation group. The fact that the simulated effect on DMIactual was detected based only on accelerations is highly significant, and models such as AEN could be substantially improved if they were derived from a greater sample size and included different physiological stages in dairy cows. Copyright © 2019 American Dairy Science Association. Published by Elsevier Inc. All rights reserved.